ViewVC Help

View File | Revision Log | Show Annotations | Download File | Root Listing

root/jsr166/jsr166/src/jsr166y/ForkJoinPool.java

Comparing jsr166/src/jsr166y/ForkJoinPool.java (file contents):
Revision 1.110 by jsr166, Fri Dec 23 00:58:29 2011 UTC vs.
Revision 1.111 by dl, Thu Jan 26 00:08:13 2012 UTC

#	Line 5 | Line 5
5		*/
6
7		package jsr166y;
8	–
8		import java.util.ArrayList;
9		import java.util.Arrays;
10		import java.util.Collection;
#	Line 20 | Line 19 | import java.util.concurrent.RejectedExec
19		import java.util.concurrent.RunnableFuture;
20		import java.util.concurrent.TimeUnit;
21		import java.util.concurrent.atomic.AtomicInteger;
22	<	import java.util.concurrent.locks.LockSupport;
22	>	import java.util.concurrent.atomic.AtomicLong;
23		import java.util.concurrent.locks.ReentrantLock;
24		import java.util.concurrent.locks.Condition;
25
#	Line 33 | Line 32 | import java.util.concurrent.locks.Condit
32		* <p>A {@code ForkJoinPool} differs from other kinds of {@link
33		* ExecutorService} mainly by virtue of employing
34		* <em>work-stealing</em>: all threads in the pool attempt to find and
35	<	* execute subtasks created by other active tasks (eventually blocking
36	<	* waiting for work if none exist). This enables efficient processing
37	<	* when most tasks spawn other subtasks (as do most {@code
38	<	* ForkJoinTask}s). When setting <em>asyncMode</em> to true in
39	<	* constructors, {@code ForkJoinPool}s may also be appropriate for use
40	<	* with event-style tasks that are never joined.
35	>	* execute tasks submitted to the pool and/or created by other active
36	>	* tasks (eventually blocking waiting for work if none exist). This
37	>	* enables efficient processing when most tasks spawn other subtasks
38	>	* (as do most {@code ForkJoinTask}s), as well as when many small
39	>	* tasks are submitted to the pool from external clients.  Especially
40	>	* when setting <em>asyncMode</em> to true in constructors, {@code
41	>	* ForkJoinPool}s may also be appropriate for use with event-style
42	>	* tasks that are never joined.
43		*
44		* <p>A {@code ForkJoinPool} is constructed with a given target
45		* parallelism level; by default, equal to the number of available
#	Line 59 | Line 60 | import java.util.concurrent.locks.Condit
60		*
61		* <p> As is the case with other ExecutorServices, there are three
62		* main task execution methods summarized in the following
63	<	* table. These are designed to be used by clients not already engaged
64	<	* in fork/join computations in the current pool.  The main forms of
65	<	* these methods accept instances of {@code ForkJoinTask}, but
66	<	* overloaded forms also allow mixed execution of plain {@code
67	<	* Runnable}- or {@code Callable}- based activities as well.  However,
68	<	* tasks that are already executing in a pool should normally
69	<	* <em>NOT</em> use these pool execution methods, but instead use the
70	<	* within-computation forms listed in the table.
63	>	* table. These are designed to be used primarily by clients not
64	>	* already engaged in fork/join computations in the current pool.  The
65	>	* main forms of these methods accept instances of {@code
66	>	* ForkJoinTask}, but overloaded forms also allow mixed execution of
67	>	* plain {@code Runnable}- or {@code Callable}- based activities as
68	>	* well.  However, tasks that are already executing in a pool should
69	>	* normally instead use the within-computation forms listed in the
70	>	* table unless using async event-style tasks that are not usually
71	>	* joined, in which case there is little difference among choice of
72	>	* methods.
73		*
74		* <table BORDER CELLPADDING=3 CELLSPACING=1>
75		*  <tr>
#	Line 125 | Line 128 | public class ForkJoinPool extends Abstra
128		/*
129		* Implementation Overview
130		*
131	<	* This class provides the central bookkeeping and control for a
132	<	* set of worker threads: Submissions from non-FJ threads enter
133	<	* into a submission queue. Workers take these tasks and typically
134	<	* split them into subtasks that may be stolen by other workers.
135	<	* Preference rules give first priority to processing tasks from
136	<	* their own queues (LIFO or FIFO, depending on mode), then to
137	<	* randomized FIFO steals of tasks in other worker queues, and
138	<	* lastly to new submissions.
131	>	* This class and its nested classes provide the main
132	>	* functionality and control for a set of worker threads:
133	>	* Submissions from non-FJ threads enter into submission
134	>	* queues. Workers take these tasks and typically split them into
135	>	* subtasks that may be stolen by other workers.  Preference rules
136	>	* give first priority to processing tasks from their own queues
137	>	* (LIFO or FIFO, depending on mode), then to randomized FIFO
138	>	* steals of tasks in other queues.
139	>	*
140	>	* WorkQueues.
141	>	* ==========
142	>	*
143	>	* Most operations occur within work-stealing queues (in nested
144	>	* class WorkQueue).  These are special forms of Deques that
145	>	* support only three of the four possible end-operations -- push,
146	>	* pop, and poll (aka steal), under the further constraints that
147	>	* push and pop are called only from the owning thread (or, as
148	>	* extended here, under a lock), while poll may be called from
149	>	* other threads.  (If you are unfamiliar with them, you probably
150	>	* want to read Herlihy and Shavit's book "The Art of
151	>	* Multiprocessor programming", chapter 16 describing these in
152	>	* more detail before proceeding.)  The main work-stealing queue
153	>	* design is roughly similar to those in the papers "Dynamic
154	>	* Circular Work-Stealing Deque" by Chase and Lev, SPAA 2005
155	>	* (http://research.sun.com/scalable/pubs/index.html) and
156	>	* "Idempotent work stealing" by Michael, Saraswat, and Vechev,
157	>	* PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186).
158	>	* The main differences ultimately stem from gc requirements that
159	>	* we null out taken slots as soon as we can, to maintain as small
160	>	* a footprint as possible even in programs generating huge
161	>	* numbers of tasks. To accomplish this, we shift the CAS
162	>	* arbitrating pop vs poll (steal) from being on the indices
163	>	* ("base" and "top") to the slots themselves.  So, both a
164	>	* successful pop and poll mainly entail a CAS of a slot from
165	>	* non-null to null.  Because we rely on CASes of references, we
166	>	* do not need tag bits on base or top.  They are simple ints as
167	>	* used in any circular array-based queue (see for example
168	>	* ArrayDeque).  Updates to the indices must still be ordered in a
169	>	* way that guarantees that top == base means the queue is empty,
170	>	* but otherwise may err on the side of possibly making the queue
171	>	* appear nonempty when a push, pop, or poll have not fully
172	>	* committed. Note that this means that the poll operation,
173	>	* considered individually, is not wait-free. One thief cannot
174	>	* successfully continue until another in-progress one (or, if
175	>	* previously empty, a push) completes.  However, in the
176	>	* aggregate, we ensure at least probabilistic non-blockingness.
177	>	* If an attempted steal fails, a thief always chooses a different
178	>	* random victim target to try next. So, in order for one thief to
179	>	* progress, it suffices for any in-progress poll or new push on
180	>	* any empty queue to complete.
181	>	*
182	>	* This approach also enables support a user mode in which local
183	>	* task processing is in FIFO, not LIFO order, simply by using
184	>	* poll rather than pop.  This can be useful in message-passing
185	>	* frameworks in which tasks are never joined.  However neither
186	>	* mode considers affinities, loads, cache localities, etc, so
187	>	* rarely provide the best possible performance on a given
188	>	* machine, but portably provide good throughput by averaging over
189	>	* these factors.  (Further, even if we did try to use such
190	>	* information, we do not usually have a basis for exploiting
191	>	* it. For example, some sets of tasks profit from cache
192	>	* affinities, but others are harmed by cache pollution effects.)
193	>	*
194	>	* WorkQueues are also used in a similar way for tasks submitted
195	>	* to the pool. We cannot mix these tasks in the same queues used
196	>	* for work-stealing (this would contaminate lifo/fifo
197	>	* processing). Instead, we loosely associate (via hashing)
198	>	* submission queues with submitting threads, and randomly scan
199	>	* these queues as well when looking for work. In essence,
200	>	* submitters act like workers except that they never take tasks,
201	>	* and they are multiplexed on to a finite number of shared work
202	>	* queues. However, classes are set up so that future extensions
203	>	* could allow submitters to optionally help perform tasks as
204	>	* well. Pool submissions from internal workers are also allowed,
205	>	* but use randomized rather than thread-hashed queue indices to
206	>	* avoid imbalance.  Insertion of tasks in shared mode requires a
207	>	* lock (mainly to protect in the case of resizing) but we use
208	>	* only a simple spinlock (using bits in field runState), because
209	>	* submitters encountering a busy queue try others so never block.
210	>	*
211	>	* Management.
212	>	* ==========
213		*
214		* The main throughput advantages of work-stealing stem from
215		* decentralized control -- workers mostly take tasks from
216		* themselves or each other. We cannot negate this in the
217		* implementation of other management responsibilities. The main
218		* tactic for avoiding bottlenecks is packing nearly all
219	<	* essentially atomic control state into a single 64bit volatile
220	<	* variable ("ctl"). This variable is read on the order of 10-100
221	<	* times as often as it is modified (always via CAS). (There is
222	<	* some additional control state, for example variable "shutdown"
223	<	* for which we can cope with uncoordinated updates.)  This
224	<	* streamlines synchronization and control at the expense of messy
225	<	* constructions needed to repack status bits upon updates.
226	<	* Updates tend not to contend with each other except during
227	<	* bursts while submitted tasks begin or end.  In some cases when
228	<	* they do contend, threads can instead do something else
229	<	* (usually, scan for tasks) until contention subsides.
230	<	*
231	<	* To enable packing, we restrict maximum parallelism to (1<<15)-1
232	<	* (which is far in excess of normal operating range) to allow
233	<	* ids, counts, and their negations (used for thresholding) to fit
234	<	* into 16bit fields.
235	<	*
236	<	* Recording Workers.  Workers are recorded in the "workers" array
237	<	* that is created upon pool construction and expanded if (rarely)
238	<	* necessary.  This is an array as opposed to some other data
239	<	* structure to support index-based random steals by workers.
240	<	* Updates to the array recording new workers and unrecording
241	<	* terminated ones are protected from each other by a seqLock
242	<	* (scanGuard) but the array is otherwise concurrently readable,
166	<	* and accessed directly by workers. To simplify index-based
219	>	* essentially atomic control state into two volatile variables
220	>	* that are by far most often read (not written) as status and
221	>	* consistency checks
222	>	*
223	>	* Field "ctl" contains 64 bits holding all the information needed
224	>	* to atomically decide to add, inactivate, enqueue (on an event
225	>	* queue), dequeue, and/or re-activate workers.  To enable this
226	>	* packing, we restrict maximum parallelism to (1<<15)-1 (which is
227	>	* far in excess of normal operating range) to allow ids, counts,
228	>	* and their negations (used for thresholding) to fit into 16bit
229	>	* fields.
230	>	*
231	>	* Field "runState" contains 32 bits needed to register and
232	>	* deregister WorkQueues, as well as to enable shutdown. It is
233	>	* only modified under a lock (normally briefly held, but
234	>	* occasionally protecting allocations and resizings) but even
235	>	* when locked remains available to check consistency.
236	>	*
237	>	* Recording WorkQueues.  WorkQueues are recorded in the
238	>	* "workQueues" array that is created upon pool construction and
239	>	* expanded if necessary.  Updates to the array while recording
240	>	* new workers and unrecording terminated ones are protected from
241	>	* each other by a lock but the array is otherwise concurrently
242	>	* readable, and accessed directly.  To simplify index-based
243		* operations, the array size is always a power of two, and all
244	<	* readers must tolerate null slots. To avoid flailing during
245	<	* start-up, the array is presized to hold twice #parallelism
246	<	* workers (which is unlikely to need further resizing during
247	<	* execution). But to avoid dealing with so many null slots,
248	<	* variable scanGuard includes a mask for the nearest power of two
249	<	* that contains all current workers.  All worker thread creation
250	<	* is on-demand, triggered by task submissions, replacement of
251	<	* terminated workers, and/or compensation for blocked
252	<	* workers. However, all other support code is set up to work with
253	<	* other policies.  To ensure that we do not hold on to worker
254	<	* references that would prevent GC, ALL accesses to workers are
255	<	* via indices into the workers array (which is one source of some
256	<	* of the messy code constructions here). In essence, the workers
257	<	* array serves as a weak reference mechanism. Thus for example
258	<	* the wait queue field of ctl stores worker indices, not worker
259	<	* references.  Access to the workers in associated methods (for
260	<	* example signalWork) must both index-check and null-check the
261	<	* IDs. All such accesses ignore bad IDs by returning out early
262	<	* from what they are doing, since this can only be associated
263	<	* with termination, in which case it is OK to give up.
264	<	*
265	<	* All uses of the workers array, as well as queue arrays, check
266	<	* that the array is non-null (even if previously non-null). This
267	<	* allows nulling during termination, which is currently not
268	<	* necessary, but remains an option for resource-revocation-based
269	<	* shutdown schemes.
244	>	* readers must tolerate null slots. Shared (submission) queues
245	>	* are at even indices, worker queues at odd indices. Grouping
246	>	* them together in this way simplifies and speeds up task
247	>	* scanning. To avoid flailing during start-up, the array is
248	>	* presized to hold twice #parallelism workers (which is unlikely
249	>	* to need further resizing during execution). But to avoid
250	>	* dealing with so many null slots, variable runState includes a
251	>	* mask for the nearest power of two that contains all current
252	>	* workers.  All worker thread creation is on-demand, triggered by
253	>	* task submissions, replacement of terminated workers, and/or
254	>	* compensation for blocked workers. However, all other support
255	>	* code is set up to work with other policies.  To ensure that we
256	>	* do not hold on to worker references that would prevent GC, ALL
257	>	* accesses to workQueues are via indices into the workQueues
258	>	* array (which is one source of some of the messy code
259	>	* constructions here). In essence, the workQueues array serves as
260	>	* a weak reference mechanism. Thus for example the wait queue
261	>	* field of ctl stores indices, not references.  Access to the
262	>	* workQueues in associated methods (for example signalWork) must
263	>	* both index-check and null-check the IDs. All such accesses
264	>	* ignore bad IDs by returning out early from what they are doing,
265	>	* since this can only be associated with termination, in which
266	>	* case it is OK to give up.
267	>	*
268	>	* All uses of the workQueues array check that it is non-null
269	>	* (even if previously non-null). This allows nulling during
270	>	* termination, which is currently not necessary, but remains an
271	>	* option for resource-revocation-based shutdown schemes. It also
272	>	* helps reduce JIT issuance of uncommon-trap code, which tends to
273	>	* unnecessarily complicate control flow in some methods.
274		*
275	<	* Wait Queuing. Unlike HPC work-stealing frameworks, we cannot
275	>	* Event Queuing. Unlike HPC work-stealing frameworks, we cannot
276		* let workers spin indefinitely scanning for tasks when none can
277		* be found immediately, and we cannot start/resume workers unless
278		* there appear to be tasks available.  On the other hand, we must
279		* quickly prod them into action when new tasks are submitted or
280	<	* generated.  We park/unpark workers after placing in an event
281	<	* wait queue when they cannot find work. This "queue" is actually
282	<	* a simple Treiber stack, headed by the "id" field of ctl, plus a
283	<	* 15bit counter value to both wake up waiters (by advancing their
284	<	* count) and avoid ABA effects. Successors are held in worker
285	<	* field "nextWait".  Queuing deals with several intrinsic races,
286	<	* mainly that a task-producing thread can miss seeing (and
280	>	* generated. In many usages, ramp-up time to activate workers is
281	>	* the main limiting factor in overall performance (this is
282	>	* compounded at program start-up by JIT compilation and
283	>	* allocation). So we try to streamline this as much as possible.
284	>	* We park/unpark workers after placing in an event wait queue
285	>	* when they cannot find work. This "queue" is actually a simple
286	>	* Treiber stack, headed by the "id" field of ctl, plus a 15bit
287	>	* counter value (that reflects the number of times a worker has
288	>	* been inactivated) to avoid ABA effects (we need only as many
289	>	* version numbers as worker threads). Successors are held in
290	>	* field WorkQueue.nextWait.  Queuing deals with several intrinsic
291	>	* races, mainly that a task-producing thread can miss seeing (and
292		* signalling) another thread that gave up looking for work but
293		* has not yet entered the wait queue. We solve this by requiring
294	<	* a full sweep of all workers both before (in scan()) and after
295	<	* (in tryAwaitWork()) a newly waiting worker is added to the wait
296	<	* queue. During a rescan, the worker might release some other
297	<	* queued worker rather than itself, which has the same net
298	<	* effect. Because enqueued workers may actually be rescanning
299	<	* rather than waiting, we set and clear the "parked" field of
300	<	* ForkJoinWorkerThread to reduce unnecessary calls to unpark.
301	<	* (Use of the parked field requires a secondary recheck to avoid
302	<	* missed signals.)
294	>	* a full sweep of all workers (via repeated calls to method
295	>	* scan()) both before and after a newly waiting worker is added
296	>	* to the wait queue. During a rescan, the worker might release
297	>	* some other queued worker rather than itself, which has the same
298	>	* net effect. Because enqueued workers may actually be rescanning
299	>	* rather than waiting, we set and clear the "parker" field of
300	>	* Workqueues to reduce unnecessary calls to unpark.  (this
301	>	* requires a secondary recheck to avoid missed signals.)  Note
302	>	* the unusual conventions about Thread.interrupts surrounding
303	>	* parking and other blocking: Because interrupts are used solely
304	>	* to alert threads to check termination, which is checked anyway
305	>	* upon blocking, we clear status (using Thread.interrupted)
306	>	* before any call to park, so that park does not immediately
307	>	* return due to status being set via some other unrelated call to
308	>	* interrupt in user code.
309		*
310		* Signalling.  We create or wake up workers only when there
311		* appears to be at least one task they might be able to find and
312		* execute.  When a submission is added or another worker adds a
313	<	* task to a queue that previously had two or fewer tasks, they
313	>	* task to a queue that previously had fewer than two tasks, they
314		* signal waiting workers (or trigger creation of new ones if
315		* fewer than the given parallelism level -- see signalWork).
316	<	* These primary signals are buttressed by signals during rescans
317	<	* as well as those performed when a worker steals a task and
318	<	* notices that there are more tasks too; together these cover the
319	<	* signals needed in cases when more than two tasks are pushed
229	<	* but untaken.
316	>	* These primary signals are buttressed by signals during rescans;
317	>	* together these cover the signals needed in cases when more
318	>	* tasks are pushed but untaken, and improve performance compared
319	>	* to having one thread wake up all workers.
320		*
321		* Trimming workers. To release resources after periods of lack of
322		* use, a worker starting to wait when the pool is quiescent will
#	Line 234 | Line 324 | public class ForkJoinPool extends Abstra
324		* SHRINK_RATE nanosecs. This will slowly propagate, eventually
325		* terminating all workers after long periods of non-use.
326		*
327	<	* Submissions. External submissions are maintained in an
328	<	* array-based queue that is structured identically to
329	<	* ForkJoinWorkerThread queues except for the use of
330	<	* submissionLock in method addSubmission. Unlike the case for
331	<	* worker queues, multiple external threads can add new
332	<	* submissions, so adding requires a lock.
333	<	*
334	<	* Compensation. Beyond work-stealing support and lifecycle
335	<	* control, the main responsibility of this framework is to take
336	<	* actions when one worker is waiting to join a task stolen (or
337	<	* always held by) another.  Because we are multiplexing many
338	<	* tasks on to a pool of workers, we can't just let them block (as
339	<	* in Thread.join).  We also cannot just reassign the joiner's
340	<	* run-time stack with another and replace it later, which would
341	<	* be a form of "continuation", that even if possible is not
342	<	* necessarily a good idea since we sometimes need both an
343	<	* unblocked task and its continuation to progress. Instead we
344	<	* combine two tactics:
327	>	* Shutdown and Termination. A call to shutdownNow atomically sets
328	>	* a runState bit and then (non-atomically) sets each workers
329	>	* runState status, cancels all unprocessed tasks, and wakes up
330	>	* all waiting workers.  Detecting whether termination should
331	>	* commence after a non-abrupt shutdown() call requires more work
332	>	* and bookkeeping. We need consensus about quiescence (i.e., that
333	>	* there is no more work). The active count provides a primary
334	>	* indication but non-abrupt shutdown still requires a rechecking
335	>	* scan for any workers that are inactive but not queued.
336	>	*
337	>	* Joining Tasks.
338	>	* ==============
339	>	*
340	>	* Any of several actions may be taken when one worker is waiting
341	>	* to join a task stolen (or always held by) another.  Because we
342	>	* are multiplexing many tasks on to a pool of workers, we can't
343	>	* just let them block (as in Thread.join).  We also cannot just
344	>	* reassign the joiner's run-time stack with another and replace
345	>	* it later, which would be a form of "continuation", that even if
346	>	* possible is not necessarily a good idea since we sometimes need
347	>	* both an unblocked task and its continuation to
348	>	* progress. Instead we combine two tactics:
349		*
350		*   Helping: Arranging for the joiner to execute some task that it
351	<	*      would be running if the steal had not occurred.  Method
258	<	*      ForkJoinWorkerThread.joinTask tracks joining->stealing
259	<	*      links to try to find such a task.
351	>	*      would be running if the steal had not occurred.
352		*
353		*   Compensating: Unless there are already enough live threads,
354	<	*      method tryPreBlock() may create or re-activate a spare
355	<	*      thread to compensate for blocked joiners until they
356	<	*      unblock.
354	>	*      method tryCompensate() may create or re-activate a spare
355	>	*      thread to compensate for blocked joiners until they unblock.
356	>	*
357	>	* A third form (implemented in tryRemoveAndExec and
358	>	* tryPollForAndExec) amounts to helping a hypothetical
359	>	* compensator: If we can readily tell that a possible action of a
360	>	* compensator is to steal and execute the task being joined, the
361	>	* joining thread can do so directly, without the need for a
362	>	* compensation thread (although at the expense of larger run-time
363	>	* stacks, but the tradeoff is typically worthwhile).
364		*
365		* The ManagedBlocker extension API can't use helping so relies
366		* only on compensation in method awaitBlocker.
367		*
368	+	* The algorithm in tryHelpStealer entails a form of "linear"
369	+	* helping: Each worker records (in field currentSteal) the most
370	+	* recent task it stole from some other worker. Plus, it records
371	+	* (in field currentJoin) the task it is currently actively
372	+	* joining. Method tryHelpStealer uses these markers to try to
373	+	* find a worker to help (i.e., steal back a task from and execute
374	+	* it) that could hasten completion of the actively joined task.
375	+	* In essence, the joiner executes a task that would be on its own
376	+	* local deque had the to-be-joined task not been stolen. This may
377	+	* be seen as a conservative variant of the approach in Wagner &
378	+	* Calder "Leapfrogging: a portable technique for implementing
379	+	* efficient futures" SIGPLAN Notices, 1993
380	+	* (http://portal.acm.org/citation.cfm?id=155354). It differs in
381	+	* that: (1) We only maintain dependency links across workers upon
382	+	* steals, rather than use per-task bookkeeping.  This sometimes
383	+	* requires a linear scan of workers array to locate stealers, but
384	+	* often doesn't because stealers leave hints (that may become
385	+	* stale/wrong) of where to locate them.  A stealHint is only a
386	+	* hint because a worker might have had multiple steals and the
387	+	* hint records only one of them (usually the most current).
388	+	* Hinting isolates cost to when it is needed, rather than adding
389	+	* to per-task overhead.  (2) It is "shallow", ignoring nesting
390	+	* and potentially cyclic mutual steals.  (3) It is intentionally
391	+	* racy: field currentJoin is updated only while actively joining,
392	+	* which means that we miss links in the chain during long-lived
393	+	* tasks, GC stalls etc (which is OK since blocking in such cases
394	+	* is usually a good idea).  (4) We bound the number of attempts
395	+	* to find work (see MAX_HELP_DEPTH) and fall back to suspending
396	+	* the worker and if necessary replacing it with another.
397	+	*
398		* It is impossible to keep exactly the target parallelism number
399		* of threads running at any given time.  Determining the
400		* existence of conservatively safe helping targets, the
401		* availability of already-created spares, and the apparent need
402	<	* to create new spares are all racy and require heuristic
403	<	* guidance, so we rely on multiple retries of each.  Currently,
404	<	* in keeping with on-demand signalling policy, we compensate only
405	<	* if blocking would leave less than one active (non-waiting,
406	<	* non-blocked) worker. Additionally, to avoid some false alarms
407	<	* due to GC, lagging counters, system activity, etc, compensated
408	<	* blocking for joins is only attempted after rechecks stabilize
409	<	* (retries are interspersed with Thread.yield, for good
410	<	* citizenship).  The variable blockedCount, incremented before
282	<	* blocking and decremented after, is sometimes needed to
283	<	* distinguish cases of waiting for work vs blocking on joins or
284	<	* other managed sync. Both cases are equivalent for most pool
285	<	* control, so we can update non-atomically. (Additionally,
286	<	* contention on blockedCount alleviates some contention on ctl).
287	<	*
288	<	* Shutdown and Termination. A call to shutdownNow atomically sets
289	<	* the ctl stop bit and then (non-atomically) sets each workers
290	<	* "terminate" status, cancels all unprocessed tasks, and wakes up
291	<	* all waiting workers.  Detecting whether termination should
292	<	* commence after a non-abrupt shutdown() call requires more work
293	<	* and bookkeeping. We need consensus about quiescence (i.e., that
294	<	* there is no more work) which is reflected in active counts so
295	<	* long as there are no current blockers, as well as possible
296	<	* re-evaluations during independent changes in blocking or
297	<	* quiescing workers.
402	>	* to create new spares are all racy, so we rely on multiple
403	>	* retries of each.  Currently, in keeping with on-demand
404	>	* signalling policy, we compensate only if blocking would leave
405	>	* less than one active (non-waiting, non-blocked) worker.
406	>	* Additionally, to avoid some false alarms due to GC, lagging
407	>	* counters, system activity, etc, compensated blocking for joins
408	>	* is only attempted after rechecks stabilize in
409	>	* ForkJoinTask.awaitJoin. (Retries are interspersed with
410	>	* Thread.yield, for good citizenship.)
411		*
412		* Style notes: There is a lot of representation-level coupling
413		* among classes ForkJoinPool, ForkJoinWorkerThread, and
414	<	* ForkJoinTask.  Most fields of ForkJoinWorkerThread maintain
415	<	* data structures managed by ForkJoinPool, so are directly
416	<	* accessed.  Conversely we allow access to "workers" array by
417	<	* workers, and direct access to ForkJoinTask.status by both
418	<	* ForkJoinPool and ForkJoinWorkerThread.  There is little point
419	<	* trying to reduce this, since any associated future changes in
420	<	* representations will need to be accompanied by algorithmic
421	<	* changes anyway. All together, these low-level implementation
422	<	* choices produce as much as a factor of 4 performance
310	<	* improvement compared to naive implementations, and enable the
311	<	* processing of billions of tasks per second, at the expense of
312	<	* some ugliness.
414	>	* ForkJoinTask.  The fields of WorkQueue maintain data structures
415	>	* managed by ForkJoinPool, so are directly accessed.  There is
416	>	* little point trying to reduce this, since any associated future
417	>	* changes in representations will need to be accompanied by
418	>	* algorithmic changes anyway. All together, these low-level
419	>	* implementation choices produce as much as a factor of 4
420	>	* performance improvement compared to naive implementations, and
421	>	* enable the processing of billions of tasks per second, at the
422	>	* expense of some ugliness.
423		*
424		* Methods signalWork() and scan() are the main bottlenecks so are
425		* especially heavily micro-optimized/mangled.  There are lots of
#	Line 326 | Line 436 | public class ForkJoinPool extends Abstra
436		* The order of declarations in this file is: (1) declarations of
437		* statics (2) fields (along with constants used when unpacking
438		* some of them), listed in an order that tends to reduce
439	<	* contention among them a bit under most JVMs.  (3) internal
440	<	* control methods (4) callbacks and other support for
441	<	* ForkJoinTask and ForkJoinWorkerThread classes, (5) exported
442	<	* methods (plus a few little helpers). (6) static block
443	<	* initializing all statics in a minimally dependent order.
439	>	* contention among them a bit under most JVMs; (3) nested
440	>	* classes; (4) internal control methods; (5) callbacks and other
441	>	* support for ForkJoinTask methods; (6) exported methods (plus a
442	>	* few little helpers); (7) static block initializing all statics
443	>	* in a minimally dependent order.
444		*/
445
446		/**
#	Line 389 | Line 499 | public class ForkJoinPool extends Abstra
499		private static final AtomicInteger poolNumberGenerator;
500
501		/**
502	<	* Generator for initial random seeds for worker victim
503	<	* selection. This is used only to create initial seeds. Random
504	<	* steals use a cheaper xorshift generator per steal attempt. We
395	<	* don't expect much contention on seedGenerator, so just use a
396	<	* plain Random.
397	<	*/
398	<	static final Random workerSeedGenerator;
399	<
400	<	/**
401	<	* Array holding all worker threads in the pool.  Initialized upon
402	<	* construction. Array size must be a power of two.  Updates and
403	<	* replacements are protected by scanGuard, but the array is
404	<	* always kept in a consistent enough state to be randomly
405	<	* accessed without locking by workers performing work-stealing,
406	<	* as well as other traversal-based methods in this class, so long
407	<	* as reads memory-acquire by first reading ctl. All readers must
408	<	* tolerate that some array slots may be null.
409	<	*/
410	<	ForkJoinWorkerThread[] workers;
411	<
412	<	/**
413	<	* Initial size for submission queue array. Must be a power of
414	<	* two.  In many applications, these always stay small so we use a
415	<	* small initial cap.
416	<	*/
417	<	private static final int INITIAL_QUEUE_CAPACITY = 8;
418	<
419	<	/**
420	<	* Maximum size for submission queue array. Must be a power of two
421	<	* less than or equal to 1 << (31 - width of array entry) to
422	<	* ensure lack of index wraparound, but is capped at a lower
423	<	* value to help users trap runaway computations.
424	<	*/
425	<	private static final int MAXIMUM_QUEUE_CAPACITY = 1 << 24; // 16M
426	<
427	<	/**
428	<	* Array serving as submission queue. Initialized upon construction.
429	<	*/
430	<	private ForkJoinTask<?>[] submissionQueue;
431	<
432	<	/**
433	<	* Lock protecting submissions array for addSubmission
434	<	*/
435	<	private final ReentrantLock submissionLock;
436	<
437	<	/**
438	<	* Condition for awaitTermination, using submissionLock for
439	<	* convenience.
440	<	*/
441	<	private final Condition termination;
442	<
443	<	/**
444	<	* Creation factory for worker threads.
445	<	*/
446	<	private final ForkJoinWorkerThreadFactory factory;
447	<
448	<	/**
449	<	* The uncaught exception handler used when any worker abruptly
450	<	* terminates.
451	<	*/
452	<	final Thread.UncaughtExceptionHandler ueh;
453	<
454	<	/**
455	<	* Prefix for assigning names to worker threads
456	<	*/
457	<	private final String workerNamePrefix;
458	<
459	<	/**
460	<	* Sum of per-thread steal counts, updated only when threads are
461	<	* idle or terminating.
462	<	*/
463	<	private volatile long stealCount;
464	<
465	<	/**
466	<	* Main pool control -- a long packed with:
502	>	* Bits and masks for control variables
503	>	*
504	>	* Field ctl is a long packed with:
505		* AC: Number of active running workers minus target parallelism (16 bits)
506		* TC: Number of total workers minus target parallelism (16 bits)
507		* ST: true if pool is terminating (1 bit)
508		* EC: the wait count of top waiting thread (15 bits)
509	<	* ID: ~poolIndex of top of Treiber stack of waiting threads (16 bits)
509	>	* ID: ~(poolIndex >>> 1) of top of Treiber stack of waiters (16 bits)
510		*
511		* When convenient, we can extract the upper 32 bits of counts and
512		* the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e =
#	Line 482 | Line 520 | public class ForkJoinPool extends Abstra
520		* negative, the pool is terminating.  To deal with these possibly
521		* negative fields, we use casts in and out of "short" and/or
522		* signed shifts to maintain signedness.
523	+	*
524	+	* When a thread is queued (inactivated), its eventCount field is
525	+	* negative, which is the only way to tell if a worker is
526	+	* prevented from executing tasks, even though it must continue to
527	+	* scan for them to avoid queuing races.
528	+	*
529	+	* Field runState is an int packed with:
530	+	* SHUTDOWN: true if shutdown is enabled (1 bit)
531	+	* SEQ:  a sequence number updated upon (de)registering workers (15 bits)
532	+	* MASK: mask (power of 2 - 1) covering all registered poolIndexes (16 bits)
533	+	*
534	+	* The combination of mask and sequence number enables simple
535	+	* consistency checks: Staleness of read-only operations on the
536	+	* workers and queues arrays can be checked by comparing runState
537	+	* before vs after the reads. The low 16 bits (i.e, anding with
538	+	* SMASK) hold (the smallest power of two covering all worker
539	+	* indices, minus one.  The mask for queues (vs workers) is twice
540	+	* this value plus 1.
541		*/
486	–	volatile long ctl;
542
543		// bit positions/shifts for fields
544		private static final int  AC_SHIFT   = 48;
#	Line 515 | Line 570 | public class ForkJoinPool extends Abstra
570		private static final int  UTC_UNIT   = 1 << UTC_SHIFT;
571
572		// masks and units for dealing with e = (int)ctl
573	<	private static final int  E_MASK     = 0x7fffffff; // no STOP_BIT
574	<	private static final int  EC_UNIT    = 1 << EC_SHIFT;
573	>	private static final int E_MASK      = 0x7fffffff; // no STOP_BIT
574	>	private static final int E_SEQ       = 1 << EC_SHIFT;
575
576	<	/**
577	<	* The target parallelism level.
578	<	*/
579	<	final int parallelism;
576	>	// runState bits
577	>	private static final int SHUTDOWN    = 1 << 31;
578	>	private static final int RS_SEQ      = 1 << 16;
579	>	private static final int RS_SEQ_MASK = 0x7fff0000;
580	>
581	>	// access mode for WorkQueue
582	>	static final int LIFO_QUEUE          =  0;
583	>	static final int FIFO_QUEUE          =  1;
584	>	static final int SHARED_QUEUE        = -1;
585
586		/**
587	<	* Index (mod submission queue length) of next element to take
588	<	* from submission queue. Usage is identical to that for
589	<	* per-worker queues -- see ForkJoinWorkerThread internal
590	<	* documentation.
587	>	* The wakeup interval (in nanoseconds) for a worker waiting for a
588	>	* task when the pool is quiescent to instead try to shrink the
589	>	* number of workers.  The exact value does not matter too
590	>	* much. It must be short enough to release resources during
591	>	* sustained periods of idleness, but not so short that threads
592	>	* are continually re-created.
593		*/
594	<	volatile int queueBase;
594	>	private static final long SHRINK_RATE =
595	>	4L * 1000L * 1000L * 1000L; // 4 seconds
596
597		/**
598	<	* Index (mod submission queue length) of next element to add
599	<	* in submission queue. Usage is identical to that for
537	<	* per-worker queues -- see ForkJoinWorkerThread internal
538	<	* documentation.
598	>	* The timeout value for attempted shrinkage, includes
599	>	* some slop to cope with system timer imprecision.
600		*/
601	<	int queueTop;
601	>	private static final long SHRINK_TIMEOUT = SHRINK_RATE - (SHRINK_RATE / 10);
602
603		/**
604	<	* True when shutdown() has been called.
604	>	* The maximum stolen->joining link depth allowed in tryHelpStealer.
605	>	* Depths for legitimate chains are unbounded, but we use a fixed
606	>	* constant to avoid (otherwise unchecked) cycles and to bound
607	>	* staleness of traversal parameters at the expense of sometimes
608	>	* blocking when we could be helping.
609		*/
610	<	volatile boolean shutdown;
610	>	private static final int MAX_HELP_DEPTH = 16;
611
612	<	/**
613	<	* True if use local fifo, not default lifo, for local polling.
614	<	* Read by, and replicated by ForkJoinWorkerThreads.
612	>	/*
613	>	* Field layout order in this class tends to matter more than one
614	>	* would like. Runtime layout order is only loosely related to
615	>	* declaration order and may differ across JVMs, but the following
616	>	* empirically works OK on current JVMs.
617	>	*/
618	>
619	>	volatile long ctl;                       // main pool control
620	>	final int parallelism;                   // parallelism level
621	>	final int localMode;                     // per-worker scheduling mode
622	>	int nextPoolIndex;                       // hint used in registerWorker
623	>	volatile int runState;                   // shutdown status, seq, and mask
624	>	WorkQueue[] workQueues;                  // main registry
625	>	final ReentrantLock lock;                // for registration
626	>	final Condition termination;             // for awaitTermination
627	>	final ForkJoinWorkerThreadFactory factory; // factory for new workers
628	>	final Thread.UncaughtExceptionHandler ueh; // per-worker UEH
629	>	final AtomicLong stealCount;             // collect counts when terminated
630	>	final AtomicInteger nextWorkerNumber;    // to create worker name string
631	>	final String workerNamePrefix;           // Prefix for assigning worker names
632	>
633	>	/**
634	>	* Queues supporting work-stealing as well as external task
635	>	* submission. See above for main rationale and algorithms.
636	>	* Implementation relies heavily on "Unsafe" intrinsics
637	>	* and selective use of "volatile":
638	>	*
639	>	* Field "base" is the index (mod array.length) of the least valid
640	>	* queue slot, which is always the next position to steal (poll)
641	>	* from if nonempty. Reads and writes require volatile orderings
642	>	* but not CAS, because updates are only performed after slot
643	>	* CASes.
644	>	*
645	>	* Field "top" is the index (mod array.length) of the next queue
646	>	* slot to push to or pop from. It is written only by owner thread
647	>	* for push, or under lock for trySharedPush, and accessed by
648	>	* other threads only after reading (volatile) base.  Both top and
649	>	* base are allowed to wrap around on overflow, but (top - base)
650	>	* (or more comonly -(base - top) to force volatile read of base
651	>	* before top) still estimates size.
652	>	*
653	>	* The array slots are read and written using the emulation of
654	>	* volatiles/atomics provided by Unsafe. Insertions must in
655	>	* general use putOrderedObject as a form of releasing store to
656	>	* ensure that all writes to the task object are ordered before
657	>	* its publication in the queue. (Although we can avoid one case
658	>	* of this when locked in trySharedPush.) All removals entail a
659	>	* CAS to null.  The array is always a power of two. To ensure
660	>	* safety of Unsafe array operations, all accesses perform
661	>	* explicit null checks and implicit bounds checks via
662	>	* power-of-two masking.
663	>	*
664	>	* In addition to basic queuing support, this class contains
665	>	* fields described elsewhere to control execution. It turns out
666	>	* to work better memory-layout-wise to include them in this
667	>	* class rather than a separate class.
668	>	*
669	>	* Performance on most platforms is very sensitive to placement of
670	>	* instances of both WorkQueues and their arrays -- we absolutely
671	>	* do not want multiple WorkQueue instances or multiple queue
672	>	* arrays sharing cache lines. (It would be best for queue objects
673	>	* and their arrays to share, but there is nothing available to
674	>	* help arrange that).  Unfortunately, because they are recorded
675	>	* in a common array, WorkQueue instances are often moved to be
676	>	* adjacent by garbage collectors. To reduce impact, we use field
677	>	* padding that works OK on common platforms; this effectively
678	>	* trades off slightly slower average field access for the sake of
679	>	* avoiding really bad worst-case access. (Until better JVM
680	>	* support is in place, this padding is dependent on transient
681	>	* properties of JVM field layout rules.)  We also take care in
682	>	* allocating and sizing and resizing the array. Non-shared queue
683	>	* arrays are initialized (via method growArray) by workers before
684	>	* use. Others are allocated on first use.
685		*/
686	<	final boolean locallyFifo;
686	>	static final class WorkQueue {
687	>	/**
688	>	* Capacity of work-stealing queue array upon initialization.
689	>	* Must be a power of two; at least 4, but set larger to
690	>	* reduce cacheline sharing among queues.
691	>	*/
692	>	static final int INITIAL_QUEUE_CAPACITY = 1 << 8;
693	>
694	>	/**
695	>	* Maximum size for queue arrays. Must be a power of two less
696	>	* than or equal to 1 << (31 - width of array entry) to ensure
697	>	* lack of wraparound of index calculations, but defined to a
698	>	* value a bit less than this to help users trap runaway
699	>	* programs before saturating systems.
700	>	*/
701	>	static final int MAXIMUM_QUEUE_CAPACITY = 1 << 26; // 64M
702	>
703	>	volatile long totalSteals; // cumulative number of steals
704	>	int seed;                  // for random scanning; initialize nonzero
705	>	volatile int eventCount;   // encoded inactivation count; < 0 if inactive
706	>	int nextWait;              // encoded record of next event waiter
707	>	int rescans;               // remaining scans until block
708	>	int nsteals;               // top-level task executions since last idle
709	>	final int mode;            // lifo, fifo, or shared
710	>	int poolIndex;             // index of this queue in pool (or 0)
711	>	int stealHint;             // index of most recent known stealer
712	>	volatile int runState;     // 1: locked, -1: terminate; else 0
713	>	volatile int base;         // index of next slot for poll
714	>	int top;                   // index of next slot for push
715	>	ForkJoinTask<?>[] array;   // the elements (initially unallocated)
716	>	final ForkJoinWorkerThread owner; // owning thread or null if shared
717	>	volatile Thread parker;    // == owner during call to park; else null
718	>	ForkJoinTask<?> currentJoin;  // task being joined in awaitJoin
719	>	ForkJoinTask<?> currentSteal; // current non-local task being executed
720	>	// Heuristic padding to ameliorate unfortunate memory placements
721	>	Object p00, p01, p02, p03, p04, p05, p06, p07, p08, p09, p0a;
722	>
723	>	WorkQueue(ForkJoinWorkerThread owner, int mode) {
724	>	this.owner = owner;
725	>	this.mode = mode;
726	>	// Place indices in the center of array (that is not yet allocated)
727	>	base = top = INITIAL_QUEUE_CAPACITY >>> 1;
728	>	}
729	>
730	>	/**
731	>	* Returns number of tasks in the queue
732	>	*/
733	>	final int queueSize() {
734	>	int n = base - top; // non-owner callers must read base first
735	>	return (n >= 0) ? 0 : -n;
736	>	}
737	>
738	>	/**
739	>	* Pushes a task. Call only by owner in unshared queues.
740	>	*
741	>	* @param task the task. Caller must ensure non-null.
742	>	* @param p, if non-null, pool to signal if necessary
743	>	* @throw RejectedExecutionException if array cannot
744	>	* be resized
745	>	*/
746	>	final void push(ForkJoinTask<?> task, ForkJoinPool p) {
747	>	boolean signal = false;
748	>	ForkJoinTask<?>[] a;
749	>	int s = top, m, n;
750	>	if ((a = array) != null) {    // ignore if queue removed
751	>	U.putOrderedObject
752	>	(a, (((m = a.length - 1) & s) << ASHIFT) + ABASE, task);
753	>	if ((n = (top = s + 1) - base) <= 2) {
754	>	if (p != null)
755	>	p.signalWork();
756	>	}
757	>	else if (n >= m)
758	>	growArray(true);
759	>	}
760	>	}
761	>
762	>	/**
763	>	* Pushes a task if lock is free and array is either big
764	>	* enough or can be resized to be big enough.
765	>	*
766	>	* @param task the task. Caller must ensure non-null.
767	>	* @return true if submitted
768	>	*/
769	>	final boolean trySharedPush(ForkJoinTask<?> task) {
770	>	boolean submitted = false;
771	>	if (runState == 0 && U.compareAndSwapInt(this, RUNSTATE, 0, 1)) {
772	>	ForkJoinTask<?>[] a = array;
773	>	int s = top, n = s - base;
774	>	try {
775	>	if ((a != null && n < a.length - 1) ||
776	>	(a = growArray(false)) != null) { // must presize
777	>	int j = (((a.length - 1) & s) << ASHIFT) + ABASE;
778	>	U.putObject(a, (long)j, task);    // don't need "ordered"
779	>	top = s + 1;
780	>	submitted = true;
781	>	}
782	>	} finally {
783	>	runState = 0;                         // unlock
784	>	}
785	>	}
786	>	return submitted;
787	>	}
788	>
789	>	/**
790	>	* Takes next task, if one exists, in FIFO order.
791	>	*/
792	>	final ForkJoinTask<?> poll() {
793	>	ForkJoinTask<?>[] a; int b, i;
794	>	while ((b = base) - top < 0 && (a = array) != null &&
795	>	(i = (a.length - 1) & b) >= 0) {
796	>	int j = (i << ASHIFT) + ABASE;
797	>	ForkJoinTask<?> t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
798	>	if (t != null && base == b &&
799	>	U.compareAndSwapObject(a, j, t, null)) {
800	>	base = b + 1;
801	>	return t;
802	>	}
803	>	}
804	>	return null;
805	>	}
806	>
807	>	/**
808	>	* Takes next task, if one exists, in LIFO order.
809	>	* Call only by owner in unshared queues.
810	>	*/
811	>	final ForkJoinTask<?> pop() {
812	>	ForkJoinTask<?> t; int m;
813	>	ForkJoinTask<?>[] a = array;
814	>	if (a != null && (m = a.length - 1) >= 0) {
815	>	for (int s; (s = top - 1) - base >= 0;) {
816	>	int j = ((m & s) << ASHIFT) + ABASE;
817	>	if ((t = (ForkJoinTask<?>)U.getObjectVolatile(a, j)) == null)
818	>	break;
819	>	if (U.compareAndSwapObject(a, j, t, null)) {
820	>	top = s;
821	>	return t;
822	>	}
823	>	}
824	>	}
825	>	return null;
826	>	}
827	>
828	>	/**
829	>	* Takes next task, if one exists, in order specified by mode.
830	>	*/
831	>	final ForkJoinTask<?> nextLocalTask() {
832	>	return mode == 0 ? pop() : poll();
833	>	}
834	>
835	>	/**
836	>	* Returns next task, if one exists, in order specified by mode.
837	>	*/
838	>	final ForkJoinTask<?> peek() {
839	>	ForkJoinTask<?>[] a = array; int m;
840	>	if (a == null || (m = a.length - 1) < 0)
841	>	return null;
842	>	int i = mode == 0 ? top - 1 : base;
843	>	int j = ((i & m) << ASHIFT) + ABASE;
844	>	return (ForkJoinTask<?>)U.getObjectVolatile(a, j);
845	>	}
846	>
847	>	/**
848	>	* Returns task at index b if b is current base of queue.
849	>	*/
850	>	final ForkJoinTask<?> pollAt(int b) {
851	>	ForkJoinTask<?>[] a; int i;
852	>	ForkJoinTask<?> task = null;
853	>	if ((a = array) != null && (i = ((a.length - 1) & b)) >= 0) {
854	>	int j = (i << ASHIFT) + ABASE;
855	>	ForkJoinTask<?> t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
856	>	if (t != null && base == b &&
857	>	U.compareAndSwapObject(a, j, t, null)) {
858	>	base = b + 1;
859	>	task = t;
860	>	}
861	>	}
862	>	return task;
863	>	}
864	>
865	>	/**
866	>	* Pops the given task only if it is at the current top.
867	>	*/
868	>	final boolean tryUnpush(ForkJoinTask<?> t) {
869	>	ForkJoinTask<?>[] a; int s;
870	>	if ((a = array) != null && (s = top) != base &&
871	>	U.compareAndSwapObject
872	>	(a, (((a.length - 1) & --s) << ASHIFT) + ABASE, t, null)) {
873	>	top = s;
874	>	return true;
875	>	}
876	>	return false;
877	>	}
878	>
879	>	/**
880	>	* Polls the given task only if it is at the current base.
881	>	*/
882	>	final boolean pollFor(ForkJoinTask<?> task) {
883	>	ForkJoinTask<?>[] a; int b, i;
884	>	if ((b = base) - top < 0 && (a = array) != null &&
885	>	(i = (a.length - 1) & b) >= 0) {
886	>	int j = (i << ASHIFT) + ABASE;
887	>	if (U.getObjectVolatile(a, j) == task && base == b &&
888	>	U.compareAndSwapObject(a, j, task, null)) {
889	>	base = b + 1;
890	>	return true;
891	>	}
892	>	}
893	>	return false;
894	>	}
895	>
896	>	/**
897	>	* If present, removes from queue and executes the given task, or
898	>	* any other cancelled task. Returns (true) immediately on any CAS
899	>	* or consistency check failure so caller can retry.
900	>	*
901	>	* @return false if no progress can be made
902	>	*/
903	>	final boolean tryRemoveAndExec(ForkJoinTask<?> task) {
904	>	boolean removed = false, empty = true, progress = true;
905	>	ForkJoinTask<?>[] a; int m, s, b, n;
906	>	if ((a = array) != null && (m = a.length - 1) >= 0 &&
907	>	(n = (s = top) - (b = base)) > 0) {
908	>	for (ForkJoinTask<?> t;;) {           // traverse from s to b
909	>	int j = ((--s & m) << ASHIFT) + ABASE;
910	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
911	>	if (t == null)                    // inconsistent length
912	>	break;
913	>	else if (t == task) {
914	>	if (s + 1 == top) {           // pop
915	>	if (!U.compareAndSwapObject(a, j, task, null))
916	>	break;
917	>	top = s;
918	>	removed = true;
919	>	}
920	>	else if (base == b)           // replace with proxy
921	>	removed = U.compareAndSwapObject(a, j, task,
922	>	new EmptyTask());
923	>	break;
924	>	}
925	>	else if (t.status >= 0)
926	>	empty = false;
927	>	else if (s + 1 == top) {          // pop and throw away
928	>	if (U.compareAndSwapObject(a, j, t, null))
929	>	top = s;
930	>	break;
931	>	}
932	>	if (--n == 0) {
933	>	if (!empty && base == b)
934	>	progress = false;
935	>	break;
936	>	}
937	>	}
938	>	}
939	>	if (removed)
940	>	task.doExec();
941	>	return progress;
942	>	}
943	>
944	>	/**
945	>	* Initializes or doubles the capacity of array. Call either
946	>	* by owner or with lock held -- it is OK for base, but not
947	>	* top, to move while resizings are in progress.
948	>	*
949	>	* @param rejectOnFailure if true, throw exception if capacity
950	>	* exceeded (relayed ultimately to user); else return null.
951	>	*/
952	>	final ForkJoinTask<?>[] growArray(boolean rejectOnFailure) {
953	>	ForkJoinTask<?>[] oldA = array;
954	>	int size = oldA != null ? oldA.length << 1 : INITIAL_QUEUE_CAPACITY;
955	>	if (size <= MAXIMUM_QUEUE_CAPACITY) {
956	>	int oldMask, t, b;
957	>	ForkJoinTask<?>[] a = array = new ForkJoinTask<?>[size];
958	>	if (oldA != null && (oldMask = oldA.length - 1) >= 0 &&
959	>	(t = top) - (b = base) > 0) {
960	>	int mask = size - 1;
961	>	do {
962	>	ForkJoinTask<?> x;
963	>	int oldj = ((b & oldMask) << ASHIFT) + ABASE;
964	>	int j    = ((b &    mask) << ASHIFT) + ABASE;
965	>	x = (ForkJoinTask<?>)U.getObjectVolatile(oldA, oldj);
966	>	if (x != null &&
967	>	U.compareAndSwapObject(oldA, oldj, x, null))
968	>	U.putObjectVolatile(a, j, x);
969	>	} while (++b != t);
970	>	}
971	>	return a;
972	>	}
973	>	else if (!rejectOnFailure)
974	>	return null;
975	>	else
976	>	throw new RejectedExecutionException("Queue capacity exceeded");
977	>	}
978	>
979	>	/**
980	>	* Removes and cancels all known tasks, ignoring any exceptions
981	>	*/
982	>	final void cancelAll() {
983	>	ForkJoinTask.cancelIgnoringExceptions(currentJoin);
984	>	ForkJoinTask.cancelIgnoringExceptions(currentSteal);
985	>	for (ForkJoinTask<?> t; (t = poll()) != null; )
986	>	ForkJoinTask.cancelIgnoringExceptions(t);
987	>	}
988	>
989	>	// Execution methods
990	>
991	>	/**
992	>	* Removes and runs tasks until empty, using local mode
993	>	* ordering.
994	>	*/
995	>	final void runLocalTasks() {
996	>	if (base - top < 0) {
997	>	for (ForkJoinTask<?> t; (t = nextLocalTask()) != null; )
998	>	t.doExec();
999	>	}
1000	>	}
1001	>
1002	>	/**
1003	>	* Executes a top-level task and any local tasks remaining
1004	>	* after execution.
1005	>	*
1006	>	* @return true unless terminating
1007	>	*/
1008	>	final boolean runTask(ForkJoinTask<?> t) {
1009	>	boolean alive = true;
1010	>	if (t != null) {
1011	>	currentSteal = t;
1012	>	t.doExec();
1013	>	runLocalTasks();
1014	>	++nsteals;
1015	>	currentSteal = null;
1016	>	}
1017	>	else if (runState < 0)            // terminating
1018	>	alive = false;
1019	>	return alive;
1020	>	}
1021	>
1022	>	/**
1023	>	* Executes a non-top-level (stolen) task
1024	>	*/
1025	>	final void runSubtask(ForkJoinTask<?> t) {
1026	>	if (t != null) {
1027	>	ForkJoinTask<?> ps = currentSteal;
1028	>	currentSteal = t;
1029	>	t.doExec();
1030	>	currentSteal = ps;
1031	>	}
1032	>	}
1033	>
1034	>	/**
1035	>	* Computes next value for random probes.  Scans don't require
1036	>	* a very high quality generator, but also not a crummy one.
1037	>	* Marsaglia xor-shift is cheap and works well enough.  Note:
1038	>	* This is manually inlined in several usages in ForkJoinPool
1039	>	* to avoid writes inside busy scan loops.
1040	>	*/
1041	>	final int nextSeed() {
1042	>	int r = seed;
1043	>	r ^= r << 13;
1044	>	r ^= r >>> 17;
1045	>	r ^= r << 5;
1046	>	return seed = r;
1047	>	}
1048	>
1049	>	// Unsafe mechanics
1050	>	private static final sun.misc.Unsafe U;
1051	>	private static final long RUNSTATE;
1052	>	private static final int ABASE;
1053	>	private static final int ASHIFT;
1054	>	static {
1055	>	int s;
1056	>	try {
1057	>	U = getUnsafe();
1058	>	Class<?> k = WorkQueue.class;
1059	>	Class<?> ak = ForkJoinTask[].class;
1060	>	RUNSTATE = U.objectFieldOffset
1061	>	(k.getDeclaredField("runState"));
1062	>	ABASE = U.arrayBaseOffset(ak);
1063	>	s = U.arrayIndexScale(ak);
1064	>	} catch (Exception e) {
1065	>	throw new Error(e);
1066	>	}
1067	>	if ((s & (s-1)) != 0)
1068	>	throw new Error("data type scale not a power of two");
1069	>	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
1070	>	}
1071	>	}
1072
1073		/**
1074	<	* The number of threads in ForkJoinWorkerThreads.helpQuiescePool.
1075	<	* When non-zero, suppresses automatic shutdown when active
1076	<	* counts become zero.
1074	>	* Class for artificial tasks that are used to replace the target
1075	>	* of local joins if they are removed from an interior queue slot
1076	>	* in WorkQueue.tryRemoveAndExec. We don't need the proxy to
1077	>	* actually do anything beyond having a unique identity.
1078		*/
1079	<	volatile int quiescerCount;
1079	>	static final class EmptyTask extends ForkJoinTask<Void> {
1080	>	EmptyTask() { status = ForkJoinTask.NORMAL; } // force done
1081	>	public Void getRawResult() { return null; }
1082	>	public void setRawResult(Void x) {}
1083	>	public boolean exec() { return true; }
1084	>	}
1085
1086		/**
1087	<	* The number of threads blocked in join.
1087	>	* Computes a hash code for the given thread. This method is
1088	>	* expected to provide higher-quality hash codes than those using
1089	>	* method hashCode().
1090		*/
1091	<	volatile int blockedCount;
1091	>	static final int hashThread(Thread t) {
1092	>	long id = (t == null)? 0L : t.getId(); // Use MurmurHash of thread id
1093	>	int h = (int)id ^ (int)(id >>> 32);
1094	>	h ^= h >>> 16;
1095	>	h *= 0x85ebca6b;
1096	>	h ^= h >>> 13;
1097	>	h *= 0xc2b2ae35;
1098	>	return h ^ (h >>> 16);
1099	>	}
1100
1101		/**
1102	<	* Counter for worker Thread names (unrelated to their poolIndex)
1102	>	* Top-level runloop for workers
1103		*/
1104	<	private volatile int nextWorkerNumber;
1104	>	final void runWorker(ForkJoinWorkerThread wt) {
1105	>	WorkQueue w = wt.workQueue;
1106	>	w.growArray(false);     // Initialize queue array and seed in this thread
1107	>	w.seed = hashThread(Thread.currentThread()) | (1 << 31); // force < 0
1108	>
1109	>	do {} while (w.runTask(scan(w)));
1110	>	}
1111	>
1112	>	// Creating, registering and deregistering workers
1113
1114		/**
1115	<	* The index for the next created worker. Accessed under scanGuard.
1115	>	* Tries to create and start a worker
1116		*/
1117	<	private int nextWorkerIndex;
1117	>	private void addWorker() {
1118	>	Throwable ex = null;
1119	>	ForkJoinWorkerThread w = null;
1120	>	try {
1121	>	if ((w = factory.newThread(this)) != null) {
1122	>	w.start();
1123	>	return;
1124	>	}
1125	>	} catch (Throwable e) {
1126	>	ex = e;
1127	>	}
1128	>	deregisterWorker(w, ex);
1129	>	}
1130
1131		/**
1132	<	* SeqLock and index masking for updates to workers array.  Locked
1133	<	* when SG_UNIT is set. Unlocking clears bit by adding
1134	<	* SG_UNIT. Staleness of read-only operations can be checked by
1135	<	* comparing scanGuard to value before the reads. The low 16 bits
580	<	* (i.e, anding with SMASK) hold (the smallest power of two
581	<	* covering all worker indices, minus one, and is used to avoid
582	<	* dealing with large numbers of null slots when the workers array
583	<	* is overallocated.
1132	>	* Callback from ForkJoinWorkerThread constructor to assign a
1133	>	* public name. This must be separate from registerWorker because
1134	>	* it is called during the "super" constructor call in
1135	>	* ForkJoinWorkerThread.
1136		*/
1137	<	volatile int scanGuard;
1138	<
1139	<	private static final int SG_UNIT = 1 << 16;
1137	>	final String nextWorkerName() {
1138	>	return workerNamePrefix.concat
1139	>	(Integer.toString(nextWorkerNumber.addAndGet(1)));
1140	>	}
1141
1142		/**
1143	<	* The wakeup interval (in nanoseconds) for a worker waiting for a
1144	<	* task when the pool is quiescent to instead try to shrink the
1145	<	* number of workers.  The exact value does not matter too
1146	<	* much. It must be short enough to release resources during
594	<	* sustained periods of idleness, but not so short that threads
595	<	* are continually re-created.
1143	>	* Callback from ForkJoinWorkerThread constructor to establish and
1144	>	* record its WorkQueue
1145	>	*
1146	>	* @param wt the worker thread
1147		*/
1148	<	private static final long SHRINK_RATE =
1149	<	4L * 1000L * 1000L * 1000L; // 4 seconds
1148	>	final void registerWorker(ForkJoinWorkerThread wt) {
1149	>	WorkQueue w = wt.workQueue;
1150	>	ReentrantLock lock = this.lock;
1151	>	lock.lock();
1152	>	try {
1153	>	int k = nextPoolIndex;
1154	>	WorkQueue[] ws = workQueues;
1155	>	if (ws != null) {                       // ignore on shutdown
1156	>	int n = ws.length;
1157	>	if (k < 0 || (k & 1) == 0 || k >= n || ws[k] != null) {
1158	>	for (k = 1; k < n && ws[k] != null; k += 2)
1159	>	;                           // workers are at odd indices
1160	>	if (k >= n)                     // resize
1161	>	workQueues = ws = Arrays.copyOf(ws, n << 1);
1162	>	}
1163	>	w.poolIndex = k;
1164	>	w.eventCount = ~(k >>> 1) & SMASK;  // Set up wait count
1165	>	ws[k] = w;                          // record worker
1166	>	nextPoolIndex = k + 2;
1167	>	int rs = runState;
1168	>	int m = rs & SMASK;                 // recalculate runState mask
1169	>	if (k > m)
1170	>	m = (m << 1) + 1;
1171	>	runState = (rs & SHUTDOWN) | ((rs + RS_SEQ) & RS_SEQ_MASK) | m;
1172	>	}
1173	>	} finally {
1174	>	lock.unlock();
1175	>	}
1176	>	}
1177
1178		/**
1179	<	* Top-level loop for worker threads: On each step: if the
1180	<	* previous step swept through all queues and found no tasks, or
1181	<	* there are excess threads, then possibly blocks. Otherwise,
1182	<	* scans for and, if found, executes a task. Returns when pool
605	<	* and/or worker terminate.
1179	>	* Final callback from terminating worker, as well as failure to
1180	>	* construct or start a worker in addWorker.  Removes record of
1181	>	* worker from array, and adjusts counts. If pool is shutting
1182	>	* down, tries to complete termination.
1183		*
1184	<	* @param w the worker
1184	>	* @param wt the worker thread or null if addWorker failed
1185	>	* @param ex the exception causing failure, or null if none
1186		*/
1187	<	final void work(ForkJoinWorkerThread w) {
1188	<	boolean swept = false;                // true on empty scans
1189	<	long c;
1190	<	while (!w.terminate && (int)(c = ctl) >= 0) {
1191	<	int a;                            // active count
1192	<	if (!swept && (a = (int)(c >> AC_SHIFT)) <= 0)
1193	<	swept = scan(w, a);
1194	<	else if (tryAwaitWork(w, c))
1195	<	swept = false;
1187	>	final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) {
1188	>	WorkQueue w = null;
1189	>	if (wt != null && (w = wt.workQueue) != null) {
1190	>	w.runState = -1;                // ensure runState is set
1191	>	stealCount.getAndAdd(w.totalSteals + w.nsteals);
1192	>	int idx = w.poolIndex;
1193	>	ReentrantLock lock = this.lock;
1194	>	lock.lock();
1195	>	try {                           // remove record from array
1196	>	WorkQueue[] ws = workQueues;
1197	>	if (ws != null && idx >= 0 && idx < ws.length && ws[idx] == w)
1198	>	ws[nextPoolIndex = idx] = null;
1199	>	} finally {
1200	>	lock.unlock();
1201	>	}
1202	>	}
1203	>
1204	>	long c;                             // adjust ctl counts
1205	>	do {} while (!U.compareAndSwapLong
1206	>	(this, CTL, c = ctl, (((c - AC_UNIT) & AC_MASK) |
1207	>	((c - TC_UNIT) & TC_MASK) |
1208	>	(c & ~(AC_MASK|TC_MASK)))));
1209	>
1210	>	if (!tryTerminate(false) && w != null) {
1211	>	w.cancelAll();                  // cancel remaining tasks
1212	>	if (w.array != null)            // suppress signal if never ran
1213	>	signalWork();               // wake up or create replacement
1214		}
1215	+
1216	+	if (ex != null)                     // rethrow
1217	+	U.throwException(ex);
1218		}
1219
1220	<	// Signalling
1220	>
1221	>	// Maintaining ctl counts
1222
1223		/**
1224	<	* Wakes up or creates a worker.
1224	>	* Increments active count; mainly called upon return from blocking
1225	>	*/
1226	>	final void incrementActiveCount() {
1227	>	long c;
1228	>	do {} while (!U.compareAndSwapLong(this, CTL, c = ctl, c + AC_UNIT));
1229	>	}
1230	>
1231	>	/**
1232	>	* Activates or creates a worker
1233		*/
1234		final void signalWork() {
1235		/*
#	Line 637 | Line 1245 | public class ForkJoinPool extends Abstra
1245		*/
1246		long c; int e, u;
1247		while ((((e = (int)(c = ctl)) | (u = (int)(c >>> 32))) &
1248	<	(INT_SIGN|SHORT_SIGN)) == (INT_SIGN|SHORT_SIGN) && e >= 0) {
1249	<	if (e > 0) {                         // release a waiting worker
1250	<	int i; ForkJoinWorkerThread w; ForkJoinWorkerThread[] ws;
1251	<	if ((ws = workers) == null ||
1252	<	(i = ~e & SMASK) >= ws.length ||
1253	<	(w = ws[i]) == null)
1248	>	(INT_SIGN|SHORT_SIGN)) == (INT_SIGN|SHORT_SIGN)) {
1249	>	WorkQueue[] ws = workQueues; int i; WorkQueue w; Thread p;
1250	>	if (e == 0) {                    // add a new worker
1251	>	if (U.compareAndSwapLong
1252	>	(this, CTL, c, (long)(((u + UTC_UNIT) & UTC_MASK) |
1253	>	((u + UAC_UNIT) & UAC_MASK)) << 32)) {
1254	>	addWorker();
1255		break;
1256	<	long nc = (((long)(w.nextWait & E_MASK)) |
1257	<	((long)(u + UAC_UNIT) << 32));
1258	<	if (w.eventCount == e &&
1259	<	UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
1260	<	w.eventCount = (e + EC_UNIT) & E_MASK;
1261	<	if (w.parked)
1262	<	UNSAFE.unpark(w);
1256	>	}
1257	>	}
1258	>	else if (e > 0 && ws != null &&
1259	>	(i = ((~e << 1) | 1) & SMASK) < ws.length &&
1260	>	(w = ws[i]) != null &&
1261	>	w.eventCount == (e | INT_SIGN)) {
1262	>	if (U.compareAndSwapLong
1263	>	(this, CTL, c, (((long)(w.nextWait & E_MASK)) |
1264	>	((long)(u + UAC_UNIT) << 32)))) {
1265	>	w.eventCount = (e + E_SEQ) & E_MASK;
1266	>	if ((p = w.parker) != null)
1267	>	U.unpark(p);         // release a waiting worker
1268		break;
1269		}
1270		}
1271	<	else if (UNSAFE.compareAndSwapLong
658	<	(this, ctlOffset, c,
659	<	(long)(((u + UTC_UNIT) & UTC_MASK) |
660	<	((u + UAC_UNIT) & UAC_MASK)) << 32)) {
661	<	addWorker();
1271	>	else
1272		break;
663	–	}
1273		}
1274		}
1275
1276		/**
1277	<	* Variant of signalWork to help release waiters on rescans.
1278	<	* Tries once to release a waiter if active count < 0.
1277	>	* Tries to decrement active count (sometimes implicitly) and
1278	>	* possibly release or create a compensating worker in preparation
1279	>	* for blocking. Fails on contention or termination.
1280		*
1281	<	* @return false if failed due to contention, else true
1281	>	* @return true if the caller can block, else should recheck and retry
1282		*/
1283	<	private boolean tryReleaseWaiter() {
1284	<	long c; int e, i; ForkJoinWorkerThread w; ForkJoinWorkerThread[] ws;
1285	<	if ((e = (int)(c = ctl)) > 0 &&
1286	<	(int)(c >> AC_SHIFT) < 0 &&
677	<	(ws = workers) != null &&
678	<	(i = ~e & SMASK) < ws.length &&
679	<	(w = ws[i]) != null) {
680	<	long nc = ((long)(w.nextWait & E_MASK) |
681	<	((c + AC_UNIT) & (AC_MASK|TC_MASK)));
682	<	if (w.eventCount != e ||
683	<	!UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc))
684	<	return false;
685	<	w.eventCount = (e + EC_UNIT) & E_MASK;
686	<	if (w.parked)
687	<	UNSAFE.unpark(w);
688	<	}
689	<	return true;
690	<	}
691	<
692	<	// Scanning for tasks
1283	>	final boolean tryCompensate() {
1284	>	WorkQueue[] ws; WorkQueue w; Thread p;
1285	>	int pc = parallelism, e, u, ac, tc, i;
1286	>	long c = ctl;
1287
1288	<	/**
1289	<	* Scans for and, if found, executes one task. Scans start at a
1290	<	* random index of workers array, and randomly select the first
1291	<	* (2*#workers)-1 probes, and then, if all empty, resort to 2
1292	<	* circular sweeps, which is necessary to check quiescence. and
1293	<	* taking a submission only if no stealable tasks were found.  The
1294	<	* steal code inside the loop is a specialized form of
1295	<	* ForkJoinWorkerThread.deqTask, followed bookkeeping to support
1296	<	* helpJoinTask and signal propagation. The code for submission
1297	<	* queues is almost identical. On each steal, the worker completes
1298	<	* not only the task, but also all local tasks that this task may
1299	<	* have generated. On detecting staleness or contention when
1300	<	* trying to take a task, this method returns without finishing
707	<	* sweep, which allows global state rechecks before retry.
708	<	*
709	<	* @param w the worker
710	<	* @param a the number of active workers
711	<	* @return true if swept all queues without finding a task
712	<	*/
713	<	private boolean scan(ForkJoinWorkerThread w, int a) {
714	<	int g = scanGuard; // mask 0 avoids useless scans if only one active
715	<	int m = (parallelism == 1 - a && blockedCount == 0) ? 0 : g & SMASK;
716	<	ForkJoinWorkerThread[] ws = workers;
717	<	if (ws == null || ws.length <= m)         // staleness check
718	<	return false;
719	<	for (int r = w.seed, k = r, j = -(m + m); j <= m + m; ++j) {
720	<	ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i;
721	<	ForkJoinWorkerThread v = ws[k & m];
722	<	if (v != null && (b = v.queueBase) != v.queueTop &&
723	<	(q = v.queue) != null && (i = (q.length - 1) & b) >= 0) {
724	<	long u = (i << ASHIFT) + ABASE;
725	<	if ((t = q[i]) != null && v.queueBase == b &&
726	<	UNSAFE.compareAndSwapObject(q, u, t, null)) {
727	<	int d = (v.queueBase = b + 1) - v.queueTop;
728	<	v.stealHint = w.poolIndex;
729	<	if (d != 0)
730	<	signalWork();             // propagate if nonempty
731	<	w.execTask(t);
1288	>	if ((e = (int)c) >= 0) {
1289	>	if ((ac = ((u = (int)(c >>> 32)) >> UAC_SHIFT)) <= 0 &&
1290	>	e != 0 && (ws = workQueues) != null &&
1291	>	(i = ((~e << 1) | 1) & SMASK) < ws.length &&
1292	>	(w = ws[i]) != null) {
1293	>	if (w.eventCount == (e | INT_SIGN) &&
1294	>	U.compareAndSwapLong
1295	>	(this, CTL, c, ((long)(w.nextWait & E_MASK) |
1296	>	(c & (AC_MASK|TC_MASK))))) {
1297	>	w.eventCount = (e + E_SEQ) & E_MASK;
1298	>	if ((p = w.parker) != null)
1299	>	U.unpark(p);
1300	>	return true;             // release an idle worker
1301		}
733	–	r ^= r << 13; r ^= r >>> 17; w.seed = r ^ (r << 5);
734	–	return false;                     // store next seed
1302		}
1303	<	else if (j < 0) {                     // xorshift
1304	<	r ^= r << 13; r ^= r >>> 17; k = r ^= r << 5;
1303	>	else if ((tc = (short)(u >>> UTC_SHIFT)) >= 0 && ac + pc > 1) {
1304	>	long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK);
1305	>	if (U.compareAndSwapLong(this, CTL, c, nc))
1306	>	return true;             // no compensation needed
1307	>	}
1308	>	else if (tc + pc < MAX_ID) {
1309	>	long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
1310	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1311	>	addWorker();
1312	>	return true;             // create replacement
1313	>	}
1314		}
739	–	else
740	–	++k;
741	–	}
742	–	if (scanGuard != g)                       // staleness check
743	–	return false;
744	–	else {                                    // try to take submission
745	–	ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i;
746	–	if ((b = queueBase) != queueTop &&
747	–	(q = submissionQueue) != null &&
748	–	(i = (q.length - 1) & b) >= 0) {
749	–	long u = (i << ASHIFT) + ABASE;
750	–	if ((t = q[i]) != null && queueBase == b &&
751	–	UNSAFE.compareAndSwapObject(q, u, t, null)) {
752	–	queueBase = b + 1;
753	–	w.execTask(t);
754	–	}
755	–	return false;
756	–	}
757	–	return true;                         // all queues empty
1315		}
1316	+	return false;
1317		}
1318
1319	+	// Submissions
1320	+
1321		/**
1322	<	* Tries to enqueue worker w in wait queue and await change in
1323	<	* worker's eventCount.  If the pool is quiescent and there is
1324	<	* more than one worker, possibly terminates worker upon exit.
1325	<	* Otherwise, before blocking, rescans queues to avoid missed
1326	<	* signals.  Upon finding work, releases at least one worker
1327	<	* (which may be the current worker). Rescans restart upon
768	<	* detected staleness or failure to release due to
769	<	* contention. Note the unusual conventions about Thread.interrupt
770	<	* here and elsewhere: Because interrupts are used solely to alert
771	<	* threads to check termination, which is checked here anyway, we
772	<	* clear status (using Thread.interrupted) before any call to
773	<	* park, so that park does not immediately return due to status
774	<	* being set via some other unrelated call to interrupt in user
775	<	* code.
776	<	*
777	<	* @param w the calling worker
778	<	* @param c the ctl value on entry
779	<	* @return true if waited or another thread was released upon enq
1322	>	* Unless shutting down, adds the given task to some submission
1323	>	* queue; using a randomly chosen queue index if the caller is a
1324	>	* ForkJoinWorkerThread, else one based on caller thread's hash
1325	>	* code. If no queue exists at the index, one is created.  If the
1326	>	* queue is busy, another is chosen by sweeping through the queues
1327	>	* array.
1328		*/
1329	<	private boolean tryAwaitWork(ForkJoinWorkerThread w, long c) {
1330	<	int v = w.eventCount;
1331	<	w.nextWait = (int)c;                      // w's successor record
1332	<	long nc = (long)(v & E_MASK) | ((c - AC_UNIT) & (AC_MASK|TC_MASK));
1333	<	if (ctl != c || !UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
1334	<	long d = ctl; // return true if lost to a deq, to force scan
1335	<	return (int)d != (int)c && (d & AC_MASK) >= (c & AC_MASK);
1336	<	}
1337	<	for (int sc = w.stealCount; sc != 0;) {   // accumulate stealCount
1338	<	long s = stealCount;
1339	<	if (UNSAFE.compareAndSwapLong(this, stealCountOffset, s, s + sc))
1340	<	sc = w.stealCount = 0;
1341	<	else if (w.eventCount != v)
1342	<	return true;                      // update next time
1343	<	}
1344	<	if ((!shutdown || !tryTerminate(false)) &&
1345	<	(int)c != 0 && parallelism + (int)(nc >> AC_SHIFT) == 0 &&
1346	<	blockedCount == 0 && quiescerCount == 0)
1347	<	idleAwaitWork(w, nc, c, v);           // quiescent
1348	<	for (boolean rescanned = false;;) {
1349	<	if (w.eventCount != v)
1350	<	return true;
1351	<	if (!rescanned) {
804	<	int g = scanGuard, m = g & SMASK;
805	<	ForkJoinWorkerThread[] ws = workers;
806	<	if (ws != null && m < ws.length) {
807	<	rescanned = true;
808	<	for (int i = 0; i <= m; ++i) {
809	<	ForkJoinWorkerThread u = ws[i];
810	<	if (u != null) {
811	<	if (u.queueBase != u.queueTop &&
812	<	!tryReleaseWaiter())
813	<	rescanned = false; // contended
814	<	if (w.eventCount != v)
815	<	return true;
816	<	}
1329	>	private void doSubmit(ForkJoinTask<?> task) {
1330	>	if (task == null)
1331	>	throw new NullPointerException();
1332	>	Thread t = Thread.currentThread();
1333	>	int r = ((t instanceof ForkJoinWorkerThread) ?
1334	>	((ForkJoinWorkerThread)t).workQueue.nextSeed() : hashThread(t));
1335	>	for (;;) {
1336	>	int rs = runState, m = rs & SMASK;
1337	>	int j = r &= (m & ~1);                      // even numbered queues
1338	>	WorkQueue[] ws = workQueues;
1339	>	if (rs < 0 || ws == null)
1340	>	throw new RejectedExecutionException(); // shutting down
1341	>	if (ws.length > m) {                        // consistency check
1342	>	for (WorkQueue q;;) {                   // circular sweep
1343	>	if (((q = ws[j]) != null ||
1344	>	(q = tryAddSharedQueue(j)) != null) &&
1345	>	q.trySharedPush(task)) {
1346	>	signalWork();
1347	>	return;
1348	>	}
1349	>	if ((j = (j + 2) & m) == r) {
1350	>	Thread.yield();                 // all queues busy
1351	>	break;
1352		}
1353		}
819	–	if (scanGuard != g ||              // stale
820	–	(queueBase != queueTop && !tryReleaseWaiter()))
821	–	rescanned = false;
822	–	if (!rescanned)
823	–	Thread.yield();                // reduce contention
824	–	else
825	–	Thread.interrupted();          // clear before park
826	–	}
827	–	else {
828	–	w.parked = true;                   // must recheck
829	–	if (w.eventCount != v) {
830	–	w.parked = false;
831	–	return true;
832	–	}
833	–	LockSupport.park(this);
834	–	rescanned = w.parked = false;
1354		}
1355		}
1356		}
1357
1358		/**
1359	<	* If inactivating worker w has caused pool to become
841	<	* quiescent, check for pool termination, and wait for event
842	<	* for up to SHRINK_RATE nanosecs (rescans are unnecessary in
843	<	* this case because quiescence reflects consensus about lack
844	<	* of work). On timeout, if ctl has not changed, terminate the
845	<	* worker. Upon its termination (see deregisterWorker), it may
846	<	* wake up another worker to possibly repeat this process.
1359	>	* Tries to add and register a new queue at the given index.
1360		*
1361	<	* @param w the calling worker
1362	<	* @param currentCtl the ctl value after enqueuing w
1363	<	* @param prevCtl the ctl value if w terminated
1364	<	* @param v the eventCount w awaits change
1365	<	*/
1366	<	private void idleAwaitWork(ForkJoinWorkerThread w, long currentCtl,
1367	<	long prevCtl, int v) {
1368	<	if (w.eventCount == v) {
1369	<	if (shutdown)
1370	<	tryTerminate(false);
1371	<	ForkJoinTask.helpExpungeStaleExceptions(); // help clean weak refs
1372	<	while (ctl == currentCtl) {
1373	<	long startTime = System.nanoTime();
1374	<	w.parked = true;
1375	<	if (w.eventCount == v)             // must recheck
1376	<	LockSupport.parkNanos(this, SHRINK_RATE);
1377	<	w.parked = false;
1378	<	if (w.eventCount != v)
1379	<	break;
1380	<	else if (System.nanoTime() - startTime <
1381	<	SHRINK_RATE - (SHRINK_RATE / 10)) // timing slop
1382	<	Thread.interrupted();          // spurious wakeup
1383	<	else if (UNSAFE.compareAndSwapLong(this, ctlOffset,
871	<	currentCtl, prevCtl)) {
872	<	w.terminate = true;            // restore previous
873	<	w.eventCount = ((int)currentCtl + EC_UNIT) & E_MASK;
874	<	break;
1361	>	* @param idx the workQueues array index to register the queue
1362	>	* @return the queue, or null if could not add because could
1363	>	* not acquire lock or idx is unusable
1364	>	*/
1365	>	private WorkQueue tryAddSharedQueue(int idx) {
1366	>	WorkQueue q = null;
1367	>	ReentrantLock lock = this.lock;
1368	>	if (idx >= 0 && (idx & 1) == 0 && !lock.isLocked()) {
1369	>	// create queue outside of lock but only if apparently free
1370	>	WorkQueue nq = new WorkQueue(null, SHARED_QUEUE);
1371	>	if (lock.tryLock()) {
1372	>	try {
1373	>	WorkQueue[] ws = workQueues;
1374	>	if (ws != null && idx < ws.length) {
1375	>	if ((q = ws[idx]) == null) {
1376	>	int rs;         // update runState seq
1377	>	ws[idx] = q = nq;
1378	>	runState = (((rs = runState) & SHUTDOWN) |
1379	>	((rs + RS_SEQ) & ~SHUTDOWN));
1380	>	}
1381	>	}
1382	>	} finally {
1383	>	lock.unlock();
1384		}
1385		}
1386		}
1387	+	return q;
1388		}
1389
1390	<	// Submissions
1390	>	// Scanning for tasks
1391
1392		/**
1393	<	* Enqueues the given task in the submissionQueue.  Same idea as
1394	<	* ForkJoinWorkerThread.pushTask except for use of submissionLock.
1395	<	*
1396	<	* @param t the task
1397	<	*/
1398	<	private void addSubmission(ForkJoinTask<?> t) {
1399	<	final ReentrantLock lock = this.submissionLock;
1400	<	lock.lock();
1401	<	try {
1402	<	ForkJoinTask<?>[] q; int s, m;
1403	<	if ((q = submissionQueue) != null) {    // ignore if queue removed
1404	<	long u = (((s = queueTop) & (m = q.length-1)) << ASHIFT)+ABASE;
1405	<	UNSAFE.putOrderedObject(q, u, t);
1406	<	queueTop = s + 1;
1407	<	if (s - queueBase == m)
1408	<	growSubmissionQueue();
1393	>	* Scans for and, if found, returns one task, else possibly
1394	>	* inactivates the worker. This method operates on single reads of
1395	>	* volatile state and is designed to be re-invoked continuously in
1396	>	* part because it returns upon detecting inconsistencies,
1397	>	* contention, or state changes that indicate possible success on
1398	>	* re-invocation.
1399	>	*
1400	>	* The scan searches for tasks across queues, randomly selecting
1401	>	* the first #queues probes, favoring steals 2:1 over submissions
1402	>	* (by exploiting even/odd indexing), and then performing a
1403	>	* circular sweep of all queues.  The scan terminates upon either
1404	>	* finding a non-empty queue, or completing a full sweep. If the
1405	>	* worker is not inactivated, it takes and returns a task from
1406	>	* this queue.  On failure to find a task, we take one of the
1407	>	* following actions, after which the caller will retry calling
1408	>	* this method unless terminated.
1409	>	*
1410	>	* * If not a complete sweep, try to release a waiting worker.  If
1411	>	* the scan terminated because the worker is inactivated, then the
1412	>	* released worker will often be the calling worker, and it can
1413	>	* succeed obtaining a task on the next call. Or maybe it is
1414	>	* another worker, but with same net effect. Releasing in other
1415	>	* cases as well ensures that we have enough workers running.
1416	>	*
1417	>	* * If the caller has run a task since the the last empty scan,
1418	>	* return (to allow rescan) if other workers are not also yet
1419	>	* enqueued.  Field WorkQueue.rescans counts down on each scan to
1420	>	* ensure eventual inactivation, and occasional calls to
1421	>	* Thread.yield to help avoid interference with more useful
1422	>	* activities on the system.
1423	>	*
1424	>	* * If pool is terminating, terminate the worker
1425	>	*
1426	>	* * If not already enqueued, try to inactivate and enqueue the
1427	>	* worker on wait queue.
1428	>	*
1429	>	* * If already enqueued and none of the above apply, either park
1430	>	* awaiting signal, or if this is the most recent waiter and pool
1431	>	* is quiescent, relay to idleAwaitWork to check for termination
1432	>	* and possibly shrink pool.
1433	>	*
1434	>	* @param w the worker (via its WorkQueue)
1435	>	* @return a task or null of none found
1436	>	*/
1437	>	private final ForkJoinTask<?> scan(WorkQueue w) {
1438	>	boolean swept = false;                 // true after full empty scan
1439	>	WorkQueue[] ws;                        // volatile read order matters
1440	>	int r = w.seed, ec = w.eventCount;     // ec is negative if inactive
1441	>	int rs = runState, m = rs & SMASK;
1442	>	if ((ws = workQueues) != null && ws.length > m) {
1443	>	ForkJoinTask<?> task = null;
1444	>	for (int k = 0, j = -2 - m; ; ++j) {
1445	>	WorkQueue q; int b;
1446	>	if (j < 0) {                    // random probes while j negative
1447	>	r ^= r << 13; r ^= r >>> 17; k = (r ^= r << 5) | (j & 1);
1448	>	}                               // worker (not submit) for odd j
1449	>	else                            // cyclic scan when j >= 0
1450	>	k += (m >>> 1) | 1;         // step by half to reduce bias
1451	>
1452	>	if ((q = ws[k & m]) != null && (b = q.base) - q.top < 0) {
1453	>	if (ec >= 0)
1454	>	task = q.pollAt(b);     // steal
1455	>	break;
1456	>	}
1457	>	else if (j > m) {
1458	>	if (rs == runState)        // staleness check
1459	>	swept = true;
1460	>	break;
1461	>	}
1462	>	}
1463	>	w.seed = r;                        // save seed for next scan
1464	>	if (task != null)
1465	>	return task;
1466	>	}
1467	>
1468	>	// Decode ctl on empty scan
1469	>	long c = ctl; int e = (int)c, a = (int)(c >> AC_SHIFT), nr, ns;
1470	>	if (!swept) {                          // try to release a waiter
1471	>	WorkQueue v; Thread p;
1472	>	if (e > 0 && a < 0 && ws != null &&
1473	>	(v = ws[((~e << 1) | 1) & m]) != null &&
1474	>	v.eventCount == (e | INT_SIGN) && U.compareAndSwapLong
1475	>	(this, CTL, c, ((long)(v.nextWait & E_MASK) |
1476	>	((c + AC_UNIT) & (AC_MASK|TC_MASK))))) {
1477	>	v.eventCount = (e + E_SEQ) & E_MASK;
1478	>	if ((p = v.parker) != null)
1479	>	U.unpark(p);
1480	>	}
1481	>	}
1482	>	else if ((nr = w.rescans) > 0) {       // continue rescanning
1483	>	int ac = a + parallelism;
1484	>	if ((w.rescans = (ac < nr) ? ac : nr - 1) > 0 && w.seed < 0 &&
1485	>	w.eventCount == ec)
1486	>	Thread.yield();                // 1 bit randomness for yield call
1487	>	}
1488	>	else if (e < 0)                        // pool is terminating
1489	>	w.runState = -1;
1490	>	else if (ec >= 0) {                    // try to enqueue
1491	>	long nc = (long)ec | ((c - AC_UNIT) & (AC_MASK|TC_MASK));
1492	>	w.nextWait = e;
1493	>	w.eventCount = ec | INT_SIGN;      // mark as inactive
1494	>	if (!U.compareAndSwapLong(this, CTL, c, nc))
1495	>	w.eventCount = ec;             // back out on CAS failure
1496	>	else if ((ns = w.nsteals) != 0) {  // set rescans if ran task
1497	>	if (a <= 0)                    // ... unless too many active
1498	>	w.rescans = a + parallelism;
1499	>	w.nsteals = 0;
1500	>	w.totalSteals += ns;
1501	>	}
1502	>	}
1503	>	else{                                  // already queued
1504	>	if (parallelism == -a)
1505	>	idleAwaitWork(w);              // quiescent
1506	>	if (w.eventCount == ec) {
1507	>	Thread.interrupted();          // clear status
1508	>	ForkJoinWorkerThread wt = w.owner;
1509	>	U.putObject(wt, PARKBLOCKER, this);
1510	>	w.parker = wt;                 // emulate LockSupport.park
1511	>	if (w.eventCount == ec)        // recheck
1512	>	U.park(false, 0L);         // block
1513	>	w.parker = null;
1514	>	U.putObject(wt, PARKBLOCKER, null);
1515		}
900	–	} finally {
901	–	lock.unlock();
1516		}
1517	<	signalWork();
1517	>	return null;
1518		}
1519
906	–	//  (pollSubmission is defined below with exported methods)
907	–
1520		/**
1521	<	* Creates or doubles submissionQueue array.
1522	<	* Basically identical to ForkJoinWorkerThread version.
1521	>	* If inactivating worker w has caused pool to become quiescent,
1522	>	* check for pool termination, and, so long as this is not the
1523	>	* only worker, wait for event for up to SHRINK_RATE nanosecs On
1524	>	* timeout, if ctl has not changed, terminate the worker, which
1525	>	* will in turn wake up another worker to possibly repeat this
1526	>	* process.
1527	>	*
1528	>	* @param w the calling worker
1529		*/
1530	<	private void growSubmissionQueue() {
1531	<	ForkJoinTask<?>[] oldQ = submissionQueue;
1532	<	int size = oldQ != null ? oldQ.length << 1 : INITIAL_QUEUE_CAPACITY;
1533	<	if (size > MAXIMUM_QUEUE_CAPACITY)
1534	<	throw new RejectedExecutionException("Queue capacity exceeded");
1535	<	if (size < INITIAL_QUEUE_CAPACITY)
1536	<	size = INITIAL_QUEUE_CAPACITY;
1537	<	ForkJoinTask<?>[] q = submissionQueue = new ForkJoinTask<?>[size];
1538	<	int mask = size - 1;
1539	<	int top = queueTop;
1540	<	int oldMask;
1541	<	if (oldQ != null && (oldMask = oldQ.length - 1) >= 0) {
1542	<	for (int b = queueBase; b != top; ++b) {
1543	<	long u = ((b & oldMask) << ASHIFT) + ABASE;
1544	<	Object x = UNSAFE.getObjectVolatile(oldQ, u);
1545	<	if (x != null && UNSAFE.compareAndSwapObject(oldQ, u, x, null))
1546	<	UNSAFE.putObjectVolatile
1547	<	(q, ((b & mask) << ASHIFT) + ABASE, x);
1530	>	private void idleAwaitWork(WorkQueue w) {
1531	>	long c; int nw, ec;
1532	>	if (!tryTerminate(false) &&
1533	>	(int)((c = ctl) >> AC_SHIFT) + parallelism == 0 &&
1534	>	(ec = w.eventCount) == ((int)c | INT_SIGN) &&
1535	>	(nw = w.nextWait) != 0) {
1536	>	long nc = ((long)(nw & E_MASK) | // ctl to restore on timeout
1537	>	((c + AC_UNIT) & AC_MASK) | (c & TC_MASK));
1538	>	ForkJoinTask.helpExpungeStaleExceptions(); // help clean
1539	>	ForkJoinWorkerThread wt = w.owner;
1540	>	while (ctl == c) {
1541	>	long startTime = System.nanoTime();
1542	>	Thread.interrupted();  // timed variant of version in scan()
1543	>	U.putObject(wt, PARKBLOCKER, this);
1544	>	w.parker = wt;
1545	>	if (ctl == c)
1546	>	U.park(false, SHRINK_RATE);
1547	>	w.parker = null;
1548	>	U.putObject(wt, PARKBLOCKER, null);
1549	>	if (ctl != c)
1550	>	break;
1551	>	if (System.nanoTime() - startTime >= SHRINK_TIMEOUT &&
1552	>	U.compareAndSwapLong(this, CTL, c, nc)) {
1553	>	w.runState = -1;          // shrink
1554	>	w.eventCount = (ec + E_SEQ) | E_MASK;
1555	>	break;
1556	>	}
1557		}
1558		}
1559		}
1560
934	–	// Blocking support
935	–
1561		/**
1562	<	* Tries to increment blockedCount, decrement active count
1563	<	* (sometimes implicitly) and possibly release or create a
1564	<	* compensating worker in preparation for blocking. Fails
1565	<	* on contention or termination.
1566	<	*
1567	<	* @return true if the caller can block, else should recheck and retry
1568	<	*/
1569	<	private boolean tryPreBlock() {
1570	<	int b = blockedCount;
1571	<	if (UNSAFE.compareAndSwapInt(this, blockedCountOffset, b, b + 1)) {
1572	<	int pc = parallelism;
1573	<	do {
1574	<	ForkJoinWorkerThread[] ws; ForkJoinWorkerThread w;
1575	<	int e, ac, tc, i;
1576	<	long c = ctl;
1577	<	int u = (int)(c >>> 32);
1578	<	if ((e = (int)c) < 0) {
1579	<	// skip -- terminating
1580	<	}
1581	<	else if ((ac = (u >> UAC_SHIFT)) <= 0 && e != 0 &&
1582	<	(ws = workers) != null &&
1583	<	(i = ~e & SMASK) < ws.length &&
1584	<	(w = ws[i]) != null) {
1585	<	long nc = ((long)(w.nextWait & E_MASK) |
1586	<	(c & (AC_MASK|TC_MASK)));
1587	<	if (w.eventCount == e &&
1588	<	UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
1589	<	w.eventCount = (e + EC_UNIT) & E_MASK;
1590	<	if (w.parked)
1591	<	UNSAFE.unpark(w);
1592	<	return true;             // release an idle worker
1562	>	* Tries to locate and execute tasks for a stealer of the given
1563	>	* task, or in turn one of its stealers, Traces currentSteal ->
1564	>	* currentJoin links looking for a thread working on a descendant
1565	>	* of the given task and with a non-empty queue to steal back and
1566	>	* execute tasks from. The first call to this method upon a
1567	>	* waiting join will often entail scanning/search, (which is OK
1568	>	* because the joiner has nothing better to do), but this method
1569	>	* leaves hints in workers to speed up subsequent calls. The
1570	>	* implementation is very branchy to cope with potential
1571	>	* inconsistencies or loops encountering chains that are stale,
1572	>	* unknown, or of length greater than MAX_HELP_DEPTH links.  All
1573	>	* of these cases are dealt with by just retrying by caller.
1574	>	*
1575	>	* @param joiner the joining worker
1576	>	* @param task the task to join
1577	>	* @return true if found or ran a task (and so is immediately retryable)
1578	>	*/
1579	>	final boolean tryHelpStealer(WorkQueue joiner, ForkJoinTask<?> task) {
1580	>	ForkJoinTask<?> subtask;    // current target
1581	>	boolean progress = false;
1582	>	int depth = 0;              // current chain depth
1583	>	int m = runState & SMASK;
1584	>	WorkQueue[] ws = workQueues;
1585	>
1586	>	if (ws != null && ws.length > m && (subtask = task).status >= 0) {
1587	>	outer:for (WorkQueue j = joiner;;) {
1588	>	// Try to find the stealer of subtask, by first using hint
1589	>	WorkQueue stealer = null;
1590	>	WorkQueue v = ws[j.stealHint & m];
1591	>	if (v != null && v.currentSteal == subtask)
1592	>	stealer = v;
1593	>	else {
1594	>	for (int i = 1; i <= m; i += 2) {
1595	>	if ((v = ws[i]) != null && v.currentSteal == subtask) {
1596	>	stealer = v;
1597	>	j.stealHint = i; // save hint
1598	>	break;
1599	>	}
1600		}
1601	+	if (stealer == null)
1602	+	break;
1603		}
1604	<	else if ((tc = (short)(u >>> UTC_SHIFT)) >= 0 && ac + pc > 1) {
1605	<	long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK);
1606	<	if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc))
1607	<	return true;             // no compensation needed
1608	<	}
1609	<	else if (tc + pc < MAX_ID) {
1610	<	long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
1611	<	if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
1612	<	addWorker();
1613	<	return true;            // create a replacement
1604	>
1605	>	for (WorkQueue q = stealer;;) { // Try to help stealer
1606	>	ForkJoinTask<?> t; int b;
1607	>	if (task.status < 0)
1608	>	break outer;
1609	>	if ((b = q.base) - q.top < 0) {
1610	>	progress = true;
1611	>	if (subtask.status < 0)
1612	>	break outer;               // stale
1613	>	if ((t = q.pollAt(b)) != null) {
1614	>	stealer.stealHint = joiner.poolIndex;
1615	>	joiner.runSubtask(t);
1616	>	}
1617	>	}
1618	>	else { // empty - try to descend to find stealer's stealer
1619	>	ForkJoinTask<?> next = stealer.currentJoin;
1620	>	if (++depth == MAX_HELP_DEPTH || subtask.status < 0 ||
1621	>	next == null || next == subtask)
1622	>	break outer;  // max depth, stale, dead-end, cyclic
1623	>	subtask = next;
1624	>	j = stealer;
1625	>	break;
1626		}
1627		}
1628	<	// try to back out on any failure and let caller retry
983	<	} while (!UNSAFE.compareAndSwapInt(this, blockedCountOffset,
984	<	b = blockedCount, b - 1));
1628	>	}
1629		}
1630	<	return false;
1630	>	return progress;
1631		}
1632
1633		/**
1634	<	* Decrements blockedCount and increments active count.
991	<	*/
992	<	private void postBlock() {
993	<	long c;
994	<	do {} while (!UNSAFE.compareAndSwapLong(this, ctlOffset,  // no mask
995	<	c = ctl, c + AC_UNIT));
996	<	int b;
997	<	do {} while (!UNSAFE.compareAndSwapInt(this, blockedCountOffset,
998	<	b = blockedCount, b - 1));
999	<	}
1000	<
1001	<	/**
1002	<	* Possibly blocks waiting for the given task to complete, or
1003	<	* cancels the task if terminating.  Fails to wait if contended.
1634	>	* If task is at base of some steal queue, steals and executes it.
1635		*
1636	<	* @param joinMe the task
1636	>	* @param joiner the joining worker
1637	>	* @param task the task
1638		*/
1639	<	final void tryAwaitJoin(ForkJoinTask<?> joinMe) {
1640	<	Thread.interrupted(); // clear interrupts before checking termination
1641	<	if (joinMe.status >= 0) {
1642	<	if (tryPreBlock()) {
1643	<	joinMe.tryAwaitDone(0L);
1644	<	postBlock();
1639	>	final void tryPollForAndExec(WorkQueue joiner, ForkJoinTask<?> task) {
1640	>	WorkQueue[] ws;
1641	>	int m = runState & SMASK;
1642	>	if ((ws = workQueues) != null && ws.length > m) {
1643	>	for (int j = 1; j <= m && task.status >= 0; j += 2) {
1644	>	WorkQueue q = ws[j];
1645	>	if (q != null && q.pollFor(task)) {
1646	>	joiner.runSubtask(task);
1647	>	break;
1648	>	}
1649		}
1014	–	else if ((ctl & STOP_BIT) != 0L)
1015	–	joinMe.cancelIgnoringExceptions();
1650		}
1651		}
1652
1653		/**
1654	<	* Possibly blocks the given worker waiting for joinMe to
1655	<	* complete or timeout.
1656	<	*
1657	<	* @param joinMe the task
1658	<	* @param millis the wait time for underlying Object.wait
1659	<	*/
1660	<	final void timedAwaitJoin(ForkJoinTask<?> joinMe, long nanos) {
1661	<	while (joinMe.status >= 0) {
1662	<	Thread.interrupted();
1663	<	if ((ctl & STOP_BIT) != 0L) {
1664	<	joinMe.cancelIgnoringExceptions();
1665	<	break;
1666	<	}
1667	<	if (tryPreBlock()) {
1668	<	long last = System.nanoTime();
1669	<	while (joinMe.status >= 0) {
1670	<	long millis = TimeUnit.NANOSECONDS.toMillis(nanos);
1037	<	if (millis <= 0)
1038	<	break;
1039	<	joinMe.tryAwaitDone(millis);
1040	<	if (joinMe.status < 0)
1041	<	break;
1042	<	if ((ctl & STOP_BIT) != 0L) {
1043	<	joinMe.cancelIgnoringExceptions();
1044	<	break;
1654	>	* Returns a non-empty steal queue, if is found during a random,
1655	>	* then cyclic scan, else null.  This method must be retried by
1656	>	* caller if, by the time it tries to use the queue, it is empty.
1657	>	*/
1658	>	private WorkQueue findNonEmptyStealQueue(WorkQueue w) {
1659	>	int r = w.seed;    // Same idea as scan(), but ignoring submissions
1660	>	for (WorkQueue[] ws;;) {
1661	>	int m = runState & SMASK;
1662	>	if ((ws = workQueues) == null)
1663	>	return null;
1664	>	if (ws.length > m) {
1665	>	WorkQueue q;
1666	>	for (int n = m << 2, k = r, j = -n;;) {
1667	>	r ^= r << 13; r ^= r >>> 17; r ^= r << 5;
1668	>	if ((q = ws[(k | 1) & m]) != null && q.base - q.top < 0) {
1669	>	w.seed = r;
1670	>	return q;
1671		}
1672	<	long now = System.nanoTime();
1673	<	nanos -= now - last;
1674	<	last = now;
1672	>	else if (j > n)
1673	>	return null;
1674	>	else
1675	>	k = (j++ < 0) ? r : k + ((m >>> 1) | 1);
1676	>
1677		}
1050	–	postBlock();
1051	–	break;
1678		}
1679		}
1680		}
1681
1682		/**
1683	<	* If necessary, compensates for blocker, and blocks.
1684	<	*/
1685	<	private void awaitBlocker(ManagedBlocker blocker)
1686	<	throws InterruptedException {
1687	<	while (!blocker.isReleasable()) {
1688	<	if (tryPreBlock()) {
1689	<	try {
1690	<	do {} while (!blocker.isReleasable() && !blocker.block());
1691	<	} finally {
1692	<	postBlock();
1683	>	* Runs tasks until {@code isQuiescent()}. We piggyback on
1684	>	* active count ctl maintenance, but rather than blocking
1685	>	* when tasks cannot be found, we rescan until all others cannot
1686	>	* find tasks either.
1687	>	*/
1688	>	final void helpQuiescePool(WorkQueue w) {
1689	>	for (boolean active = true;;) {
1690	>	w.runLocalTasks();      // exhaust local queue
1691	>	WorkQueue q = findNonEmptyStealQueue(w);
1692	>	if (q != null) {
1693	>	ForkJoinTask<?> t;
1694	>	if (!active) {      // re-establish active count
1695	>	long c;
1696	>	active = true;
1697	>	do {} while (!U.compareAndSwapLong
1698	>	(this, CTL, c = ctl, c + AC_UNIT));
1699	>	}
1700	>	if ((t = q.poll()) != null)
1701	>	w.runSubtask(t);
1702	>	}
1703	>	else {
1704	>	long c;
1705	>	if (active) {       // decrement active count without queuing
1706	>	active = false;
1707	>	do {} while (!U.compareAndSwapLong
1708	>	(this, CTL, c = ctl, c -= AC_UNIT));
1709	>	}
1710	>	else
1711	>	c = ctl;        // re-increment on exit
1712	>	if ((int)(c >> AC_SHIFT) + parallelism == 0) {
1713	>	do {} while (!U.compareAndSwapLong
1714	>	(this, CTL, c = ctl, c + AC_UNIT));
1715	>	break;
1716		}
1068	–	break;
1717		}
1718		}
1719		}
1720
1073	–	// Creating, registering and deregistering workers
1074	–
1721		/**
1722	<	* Tries to create and start a worker; minimally rolls back counts
1723	<	* on failure.
1722	>	* Gets and removes a local or stolen task for the given worker
1723	>	*
1724	>	* @return a task, if available
1725		*/
1726	<	private void addWorker() {
1727	<	Throwable ex = null;
1728	<	ForkJoinWorkerThread t = null;
1729	<	try {
1730	<	t = factory.newThread(this);
1731	<	} catch (Throwable e) {
1732	<	ex = e;
1733	<	}
1734	<	if (t == null) {  // null or exceptional factory return
1088	<	long c;       // adjust counts
1089	<	do {} while (!UNSAFE.compareAndSwapLong
1090	<	(this, ctlOffset, c = ctl,
1091	<	(((c - AC_UNIT) & AC_MASK) |
1092	<	((c - TC_UNIT) & TC_MASK) |
1093	<	(c & ~(AC_MASK|TC_MASK)))));
1094	<	// Propagate exception if originating from an external caller
1095	<	if (!tryTerminate(false) && ex != null &&
1096	<	!(Thread.currentThread() instanceof ForkJoinWorkerThread))
1097	<	UNSAFE.throwException(ex);
1726	>	final ForkJoinTask<?> nextTaskFor(WorkQueue w) {
1727	>	for (ForkJoinTask<?> t;;) {
1728	>	WorkQueue q;
1729	>	if ((t = w.nextLocalTask()) != null)
1730	>	return t;
1731	>	if ((q = findNonEmptyStealQueue(w)) == null)
1732	>	return null;
1733	>	if ((t = q.poll()) != null)
1734	>	return t;
1735		}
1099	–	else
1100	–	t.start();
1736		}
1737
1738		/**
1739	<	* Callback from ForkJoinWorkerThread constructor to assign a
1740	<	* public name
1739	>	* Returns the approximate (non-atomic) number of idle threads per
1740	>	* active thread to offset steal queue size for method
1741	>	* ForkJoinTask.getSurplusQueuedTaskCount().
1742		*/
1743	<	final String nextWorkerName() {
1744	<	for (int n;;) {
1745	<	if (UNSAFE.compareAndSwapInt(this, nextWorkerNumberOffset,
1746	<	n = nextWorkerNumber, ++n))
1747	<	return workerNamePrefix + n;
1748	<	}
1743	>	final int idlePerActive() {
1744	>	// Approximate at powers of two for small values, saturate past 4
1745	>	int p = parallelism;
1746	>	int a = p + (int)(ctl >> AC_SHIFT);
1747	>	return (a > (p >>>= 1) ? 0 :
1748	>	a > (p >>>= 1) ? 1 :
1749	>	a > (p >>>= 1) ? 2 :
1750	>	a > (p >>>= 1) ? 4 :
1751	>	8);
1752		}
1753
1754	<	/**
1116	<	* Callback from ForkJoinWorkerThread constructor to
1117	<	* determine its poolIndex and record in workers array.
1118	<	*
1119	<	* @param w the worker
1120	<	* @return the worker's pool index
1121	<	*/
1122	<	final int registerWorker(ForkJoinWorkerThread w) {
1123	<	/*
1124	<	* In the typical case, a new worker acquires the lock, uses
1125	<	* next available index and returns quickly.  Since we should
1126	<	* not block callers (ultimately from signalWork or
1127	<	* tryPreBlock) waiting for the lock needed to do this, we
1128	<	* instead help release other workers while waiting for the
1129	<	* lock.
1130	<	*/
1131	<	for (int g;;) {
1132	<	ForkJoinWorkerThread[] ws;
1133	<	if (((g = scanGuard) & SG_UNIT) == 0 &&
1134	<	UNSAFE.compareAndSwapInt(this, scanGuardOffset,
1135	<	g, g | SG_UNIT)) {
1136	<	int k = nextWorkerIndex;
1137	<	try {
1138	<	if ((ws = workers) != null) { // ignore on shutdown
1139	<	int n = ws.length;
1140	<	if (k < 0 || k >= n || ws[k] != null) {
1141	<	for (k = 0; k < n && ws[k] != null; ++k)
1142	<	;
1143	<	if (k == n)
1144	<	ws = workers = Arrays.copyOf(ws, n << 1);
1145	<	}
1146	<	ws[k] = w;
1147	<	nextWorkerIndex = k + 1;
1148	<	int m = g & SMASK;
1149	<	g = (k > m) ? ((m << 1) + 1) & SMASK : g + (SG_UNIT<<1);
1150	<	}
1151	<	} finally {
1152	<	scanGuard = g;
1153	<	}
1154	<	return k;
1155	<	}
1156	<	else if ((ws = workers) != null) { // help release others
1157	<	for (ForkJoinWorkerThread u : ws) {
1158	<	if (u != null && u.queueBase != u.queueTop) {
1159	<	if (tryReleaseWaiter())
1160	<	break;
1161	<	}
1162	<	}
1163	<	}
1164	<	}
1165	<	}
1754	>	// Termination
1755
1756		/**
1757	<	* Final callback from terminating worker.  Removes record of
1169	<	* worker from array, and adjusts counts. If pool is shutting
1170	<	* down, tries to complete termination.
1171	<	*
1172	<	* @param w the worker
1757	>	* Sets SHUTDOWN bit of runState under lock
1758		*/
1759	<	final void deregisterWorker(ForkJoinWorkerThread w, Throwable ex) {
1760	<	int idx = w.poolIndex;
1761	<	int sc = w.stealCount;
1762	<	int steps = 0;
1763	<	// Remove from array, adjust worker counts and collect steal count.
1764	<	// We can intermix failed removes or adjusts with steal updates
1180	<	do {
1181	<	long s, c;
1182	<	int g;
1183	<	if (steps == 0 && ((g = scanGuard) & SG_UNIT) == 0 &&
1184	<	UNSAFE.compareAndSwapInt(this, scanGuardOffset,
1185	<	g, g |= SG_UNIT)) {
1186	<	ForkJoinWorkerThread[] ws = workers;
1187	<	if (ws != null && idx >= 0 &&
1188	<	idx < ws.length && ws[idx] == w)
1189	<	ws[idx] = null;    // verify
1190	<	nextWorkerIndex = idx;
1191	<	scanGuard = g + SG_UNIT;
1192	<	steps = 1;
1193	<	}
1194	<	if (steps == 1 &&
1195	<	UNSAFE.compareAndSwapLong(this, ctlOffset, c = ctl,
1196	<	(((c - AC_UNIT) & AC_MASK) |
1197	<	((c - TC_UNIT) & TC_MASK) |
1198	<	(c & ~(AC_MASK|TC_MASK)))))
1199	<	steps = 2;
1200	<	if (sc != 0 &&
1201	<	UNSAFE.compareAndSwapLong(this, stealCountOffset,
1202	<	s = stealCount, s + sc))
1203	<	sc = 0;
1204	<	} while (steps != 2 || sc != 0);
1205	<	if (!tryTerminate(false)) {
1206	<	if (ex != null)   // possibly replace if died abnormally
1207	<	signalWork();
1208	<	else
1209	<	tryReleaseWaiter();
1759	>	private void enableShutdown() {
1760	>	ReentrantLock lock = this.lock;
1761	>	if (runState >= 0) {
1762	>	lock.lock();                       // don't need try/finally
1763	>	runState |= SHUTDOWN;
1764	>	lock.unlock();
1765		}
1766		}
1767
1213	–	// Shutdown and termination
1214	–
1768		/**
1769	<	* Possibly initiates and/or completes termination.
1769	>	* Possibly initiates and/or completes termination.  Upon
1770	>	* termination, cancels all queued tasks and then
1771		*
1772		* @param now if true, unconditionally terminate, else only
1773	<	* if shutdown and empty queue and no active workers
1773	>	* if no work and no active workers
1774		* @return true if now terminating or terminated
1775		*/
1776		private boolean tryTerminate(boolean now) {
1777	<	long c;
1778	<	while (((c = ctl) & STOP_BIT) == 0) {
1777	>	for (long c;;) {
1778	>	if (((c = ctl) & STOP_BIT) != 0) {      // already terminating
1779	>	if ((short)(c >>> TC_SHIFT) == -parallelism) {
1780	>	ReentrantLock lock = this.lock; // signal when no workers
1781	>	lock.lock();                    // don't need try/finally
1782	>	termination.signalAll();        // signal when 0 workers
1783	>	lock.unlock();
1784	>	}
1785	>	return true;
1786	>	}
1787		if (!now) {
1788	<	if ((int)(c >> AC_SHIFT) != -parallelism)
1788	>	if ((int)(c >> AC_SHIFT) != -parallelism || runState >= 0 ||
1789	>	hasQueuedSubmissions())
1790		return false;
1791	<	if (!shutdown || blockedCount != 0 || quiescerCount != 0 ||
1792	<	queueBase != queueTop) {
1793	<	if (ctl == c) // staleness check
1794	<	return false;
1795	<	continue;
1791	>	// Check for unqueued inactive workers. One pass suffices.
1792	>	WorkQueue[] ws = workQueues; WorkQueue w;
1793	>	if (ws != null) {
1794	>	int n = ws.length;
1795	>	for (int i = 1; i < n; i += 2) {
1796	>	if ((w = ws[i]) != null && w.eventCount >= 0)
1797	>	return false;
1798	>	}
1799		}
1800		}
1801	<	if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, c | STOP_BIT))
1801	>	if (U.compareAndSwapLong(this, CTL, c, c | STOP_BIT))
1802		startTerminating();
1803		}
1238	–	if ((short)(c >>> TC_SHIFT) == -parallelism) { // signal when 0 workers
1239	–	final ReentrantLock lock = this.submissionLock;
1240	–	lock.lock();
1241	–	try {
1242	–	termination.signalAll();
1243	–	} finally {
1244	–	lock.unlock();
1245	–	}
1246	–	}
1247	–	return true;
1804		}
1805
1806		/**
1807	<	* Runs up to three passes through workers: (0) Setting
1808	<	* termination status for each worker, followed by wakeups up to
1809	<	* queued workers; (1) helping cancel tasks; (2) interrupting
1810	<	* lagging threads (likely in external tasks, but possibly also
1811	<	* blocked in joins).  Each pass repeats previous steps because of
1812	<	* potential lagging thread creation.
1807	>	* Initiates termination: Runs three passes through workQueues:
1808	>	* (0) Setting termination status, followed by wakeups of queued
1809	>	* workers; (1) cancelling all tasks; (2) interrupting lagging
1810	>	* threads (likely in external tasks, but possibly also blocked in
1811	>	* joins).  Each pass repeats previous steps because of potential
1812	>	* lagging thread creation.
1813		*/
1814		private void startTerminating() {
1259	–	cancelSubmissions();
1815		for (int pass = 0; pass < 3; ++pass) {
1816	<	ForkJoinWorkerThread[] ws = workers;
1816	>	WorkQueue[] ws = workQueues;
1817		if (ws != null) {
1818	<	for (ForkJoinWorkerThread w : ws) {
1819	<	if (w != null) {
1820	<	w.terminate = true;
1818	>	WorkQueue w; Thread wt;
1819	>	int n = ws.length;
1820	>	for (int j = 0; j < n; ++j) {
1821	>	if ((w = ws[j]) != null) {
1822	>	w.runState = -1;
1823		if (pass > 0) {
1824	<	w.cancelTasks();
1825	<	if (pass > 1 && !w.isInterrupted()) {
1824	>	w.cancelAll();
1825	>	if (pass > 1 && (wt = w.owner) != null &&
1826	>	!wt.isInterrupted()) {
1827		try {
1828	<	w.interrupt();
1828	>	wt.interrupt();
1829		} catch (SecurityException ignore) {
1830		}
1831		}
1832		}
1833		}
1834		}
1835	<	terminateWaiters();
1836	<	}
1837	<	}
1838	<	}
1839	<
1840	<	/**
1841	<	* Polls and cancels all submissions. Called only during termination.
1842	<	*/
1843	<	private void cancelSubmissions() {
1844	<	while (queueBase != queueTop) {
1845	<	ForkJoinTask<?> task = pollSubmission();
1846	<	if (task != null) {
1847	<	try {
1290	<	task.cancel(false);
1291	<	} catch (Throwable ignore) {
1292	<	}
1293	<	}
1294	<	}
1295	<	}
1296	<
1297	<	/**
1298	<	* Tries to set the termination status of waiting workers, and
1299	<	* then wakes them up (after which they will terminate).
1300	<	*/
1301	<	private void terminateWaiters() {
1302	<	ForkJoinWorkerThread[] ws = workers;
1303	<	if (ws != null) {
1304	<	ForkJoinWorkerThread w; long c; int i, e;
1305	<	int n = ws.length;
1306	<	while ((i = ~(e = (int)(c = ctl)) & SMASK) < n &&
1307	<	(w = ws[i]) != null && w.eventCount == (e & E_MASK)) {
1308	<	if (UNSAFE.compareAndSwapLong(this, ctlOffset, c,
1309	<	(long)(w.nextWait & E_MASK) |
1310	<	((c + AC_UNIT) & AC_MASK) |
1311	<	(c & (TC_MASK|STOP_BIT)))) {
1312	<	w.terminate = true;
1313	<	w.eventCount = e + EC_UNIT;
1314	<	if (w.parked)
1315	<	UNSAFE.unpark(w);
1835	>	// Wake up workers parked on event queue
1836	>	int i, e; long c; Thread p;
1837	>	while ((i = ((~(e = (int)(c = ctl)) << 1) | 1) & SMASK) < n &&
1838	>	(w = ws[i]) != null &&
1839	>	w.eventCount == (e | INT_SIGN)) {
1840	>	long nc = ((long)(w.nextWait & E_MASK) |
1841	>	((c + AC_UNIT) & AC_MASK) |
1842	>	(c & (TC_MASK|STOP_BIT)));
1843	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1844	>	w.eventCount = (e + E_SEQ) & E_MASK;
1845	>	if ((p = w.parker) != null)
1846	>	U.unpark(p);
1847	>	}
1848		}
1849		}
1850		}
1851		}
1852
1321	–	// misc ForkJoinWorkerThread support
1322	–
1323	–	/**
1324	–	* Increments or decrements quiescerCount. Needed only to prevent
1325	–	* triggering shutdown if a worker is transiently inactive while
1326	–	* checking quiescence.
1327	–	*
1328	–	* @param delta 1 for increment, -1 for decrement
1329	–	*/
1330	–	final void addQuiescerCount(int delta) {
1331	–	int c;
1332	–	do {} while (!UNSAFE.compareAndSwapInt(this, quiescerCountOffset,
1333	–	c = quiescerCount, c + delta));
1334	–	}
1335	–
1336	–	/**
1337	–	* Directly increments or decrements active count without queuing.
1338	–	* This method is used to transiently assert inactivation while
1339	–	* checking quiescence.
1340	–	*
1341	–	* @param delta 1 for increment, -1 for decrement
1342	–	*/
1343	–	final void addActiveCount(int delta) {
1344	–	long d = (long)delta << AC_SHIFT;
1345	–	long c;
1346	–	do {} while (!UNSAFE.compareAndSwapLong(this, ctlOffset,
1347	–	c = ctl, c + d));
1348	–	}
1349	–
1350	–	/**
1351	–	* Returns the approximate (non-atomic) number of idle threads per
1352	–	* active thread.
1353	–	*/
1354	–	final int idlePerActive() {
1355	–	// Approximate at powers of two for small values, saturate past 4
1356	–	int p = parallelism;
1357	–	int a = p + (int)(ctl >> AC_SHIFT);
1358	–	return (a > (p >>>= 1) ? 0 :
1359	–	a > (p >>>= 1) ? 1 :
1360	–	a > (p >>>= 1) ? 2 :
1361	–	a > (p >>>= 1) ? 4 :
1362	–	8);
1363	–	}
1364	–
1853		// Exported methods
1854
1855		// Constructors
#	Line 1436 | Line 1924 | public class ForkJoinPool extends Abstra
1924		this.parallelism = parallelism;
1925		this.factory = factory;
1926		this.ueh = handler;
1927	<	this.locallyFifo = asyncMode;
1927	>	this.localMode = asyncMode ? FIFO_QUEUE : LIFO_QUEUE;
1928	>	this.nextPoolIndex = 1;
1929		long np = (long)(-parallelism); // offset ctl counts
1930		this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
1931	<	this.submissionQueue = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY];
1443	<	// initialize workers array with room for 2*parallelism if possible
1931	>	// initialize workQueues array with room for 2*parallelism if possible
1932		int n = parallelism << 1;
1933		if (n >= MAX_ID)
1934		n = MAX_ID;
1935		else { // See Hackers Delight, sec 3.2, where n < (1 << 16)
1936		n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8;
1937		}
1938	<	workers = new ForkJoinWorkerThread[n + 1];
1939	<	this.submissionLock = new ReentrantLock();
1940	<	this.termination = submissionLock.newCondition();
1938	>	this.workQueues = new WorkQueue[(n + 1) << 1];
1939	>	ReentrantLock lck = this.lock = new ReentrantLock();
1940	>	this.termination = lck.newCondition();
1941	>	this.stealCount = new AtomicLong();
1942	>	this.nextWorkerNumber = new AtomicInteger();
1943		StringBuilder sb = new StringBuilder("ForkJoinPool-");
1944		sb.append(poolNumberGenerator.incrementAndGet());
1945		sb.append("-worker-");
1946		this.workerNamePrefix = sb.toString();
1947	+	// Create initial submission queue
1948	+	WorkQueue sq = tryAddSharedQueue(0);
1949	+	if (sq != null)
1950	+	sq.growArray(false);
1951		}
1952
1953		// Execution methods
#	Line 1475 | Line 1969 | public class ForkJoinPool extends Abstra
1969		*         scheduled for execution
1970		*/
1971		public <T> T invoke(ForkJoinTask<T> task) {
1972	<	Thread t = Thread.currentThread();
1973	<	if (task == null)
1480	<	throw new NullPointerException();
1481	<	if (shutdown)
1482	<	throw new RejectedExecutionException();
1483	<	if ((t instanceof ForkJoinWorkerThread) &&
1484	<	((ForkJoinWorkerThread)t).pool == this)
1485	<	return task.invoke();  // bypass submit if in same pool
1486	<	else {
1487	<	addSubmission(task);
1488	<	return task.join();
1489	<	}
1490	<	}
1491	<
1492	<	/**
1493	<	* Unless terminating, forks task if within an ongoing FJ
1494	<	* computation in the current pool, else submits as external task.
1495	<	*/
1496	<	private <T> void forkOrSubmit(ForkJoinTask<T> task) {
1497	<	ForkJoinWorkerThread w;
1498	<	Thread t = Thread.currentThread();
1499	<	if (shutdown)
1500	<	throw new RejectedExecutionException();
1501	<	if ((t instanceof ForkJoinWorkerThread) &&
1502	<	(w = (ForkJoinWorkerThread)t).pool == this)
1503	<	w.pushTask(task);
1504	<	else
1505	<	addSubmission(task);
1972	>	doSubmit(task);
1973	>	return task.join();
1974		}
1975
1976		/**
#	Line 1514 | Line 1982 | public class ForkJoinPool extends Abstra
1982		*         scheduled for execution
1983		*/
1984		public void execute(ForkJoinTask<?> task) {
1985	<	if (task == null)
1518	<	throw new NullPointerException();
1519	<	forkOrSubmit(task);
1985	>	doSubmit(task);
1986		}
1987
1988		// AbstractExecutorService methods
#	Line 1534 | Line 2000 | public class ForkJoinPool extends Abstra
2000		job = (ForkJoinTask<?>) task;
2001		else
2002		job = ForkJoinTask.adapt(task, null);
2003	<	forkOrSubmit(job);
2003	>	doSubmit(job);
2004		}
2005
2006		/**
#	Line 1547 | Line 2013 | public class ForkJoinPool extends Abstra
2013		*         scheduled for execution
2014		*/
2015		public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
2016	<	if (task == null)
1551	<	throw new NullPointerException();
1552	<	forkOrSubmit(task);
2016	>	doSubmit(task);
2017		return task;
2018		}
2019
#	Line 1562 | Line 2026 | public class ForkJoinPool extends Abstra
2026		if (task == null)
2027		throw new NullPointerException();
2028		ForkJoinTask<T> job = ForkJoinTask.adapt(task);
2029	<	forkOrSubmit(job);
2029	>	doSubmit(job);
2030		return job;
2031		}
2032
#	Line 1575 | Line 2039 | public class ForkJoinPool extends Abstra
2039		if (task == null)
2040		throw new NullPointerException();
2041		ForkJoinTask<T> job = ForkJoinTask.adapt(task, result);
2042	<	forkOrSubmit(job);
2042	>	doSubmit(job);
2043		return job;
2044		}
2045
#	Line 1592 | Line 2056 | public class ForkJoinPool extends Abstra
2056		job = (ForkJoinTask<?>) task;
2057		else
2058		job = ForkJoinTask.adapt(task, null);
2059	<	forkOrSubmit(job);
2059	>	doSubmit(job);
2060		return job;
2061		}
2062
#	Line 1669 | Line 2133 | public class ForkJoinPool extends Abstra
2133		* @return {@code true} if this pool uses async mode
2134		*/
2135		public boolean getAsyncMode() {
2136	<	return locallyFifo;
2136	>	return localMode != 0;
2137		}
2138
2139		/**
#	Line 1681 | Line 2145 | public class ForkJoinPool extends Abstra
2145		* @return the number of worker threads
2146		*/
2147		public int getRunningThreadCount() {
2148	<	int r = parallelism + (int)(ctl >> AC_SHIFT);
2149	<	return (r <= 0) ? 0 : r; // suppress momentarily negative values
2148	>	int rc = 0;
2149	>	WorkQueue[] ws; WorkQueue w;
2150	>	if ((ws = workQueues) != null) {
2151	>	int n = ws.length;
2152	>	for (int i = 1; i < n; i += 2) {
2153	>	Thread.State s; ForkJoinWorkerThread wt;
2154	>	if ((w = ws[i]) != null && (wt = w.owner) != null &&
2155	>	w.eventCount >= 0 &&
2156	>	(s = wt.getState()) != Thread.State.BLOCKED &&
2157	>	s != Thread.State.WAITING &&
2158	>	s != Thread.State.TIMED_WAITING)
2159	>	++rc;
2160	>	}
2161	>	}
2162	>	return rc;
2163		}
2164
2165		/**
#	Line 1693 | Line 2170 | public class ForkJoinPool extends Abstra
2170		* @return the number of active threads
2171		*/
2172		public int getActiveThreadCount() {
2173	<	int r = parallelism + (int)(ctl >> AC_SHIFT) + blockedCount;
2173	>	int r = parallelism + (int)(ctl >> AC_SHIFT);
2174		return (r <= 0) ? 0 : r; // suppress momentarily negative values
2175		}
2176
#	Line 1709 | Line 2186 | public class ForkJoinPool extends Abstra
2186		* @return {@code true} if all threads are currently idle
2187		*/
2188		public boolean isQuiescent() {
2189	<	return parallelism + (int)(ctl >> AC_SHIFT) + blockedCount == 0;
2189	>	return (int)(ctl >> AC_SHIFT) + parallelism == 0;
2190		}
2191
2192		/**
#	Line 1724 | Line 2201 | public class ForkJoinPool extends Abstra
2201		* @return the number of steals
2202		*/
2203		public long getStealCount() {
2204	<	return stealCount;
2204	>	long count = stealCount.get();
2205	>	WorkQueue[] ws; WorkQueue w;
2206	>	if ((ws = workQueues) != null) {
2207	>	int n = ws.length;
2208	>	for (int i = 1; i < n; i += 2) {
2209	>	if ((w = ws[i]) != null)
2210	>	count += w.totalSteals;
2211	>	}
2212	>	}
2213	>	return count;
2214		}
2215
2216		/**
#	Line 1739 | Line 2225 | public class ForkJoinPool extends Abstra
2225		*/
2226		public long getQueuedTaskCount() {
2227		long count = 0;
2228	<	ForkJoinWorkerThread[] ws;
2229	<	if ((short)(ctl >>> TC_SHIFT) > -parallelism &&
2230	<	(ws = workers) != null) {
2231	<	for (ForkJoinWorkerThread w : ws)
2232	<	if (w != null)
2233	<	count -= w.queueBase - w.queueTop; // must read base first
2228	>	WorkQueue[] ws; WorkQueue w;
2229	>	if ((ws = workQueues) != null) {
2230	>	int n = ws.length;
2231	>	for (int i = 1; i < n; i += 2) {
2232	>	if ((w = ws[i]) != null)
2233	>	count += w.queueSize();
2234	>	}
2235		}
2236		return count;
2237		}
#	Line 1757 | Line 2244 | public class ForkJoinPool extends Abstra
2244		* @return the number of queued submissions
2245		*/
2246		public int getQueuedSubmissionCount() {
2247	<	return -queueBase + queueTop;
2247	>	int count = 0;
2248	>	WorkQueue[] ws; WorkQueue w;
2249	>	if ((ws = workQueues) != null) {
2250	>	int n = ws.length;
2251	>	for (int i = 0; i < n; i += 2) {
2252	>	if ((w = ws[i]) != null)
2253	>	count += w.queueSize();
2254	>	}
2255	>	}
2256	>	return count;
2257		}
2258
2259		/**
#	Line 1767 | Line 2263 | public class ForkJoinPool extends Abstra
2263		* @return {@code true} if there are any queued submissions
2264		*/
2265		public boolean hasQueuedSubmissions() {
2266	<	return queueBase != queueTop;
2266	>	WorkQueue[] ws; WorkQueue w;
2267	>	if ((ws = workQueues) != null) {
2268	>	int n = ws.length;
2269	>	for (int i = 0; i < n; i += 2) {
2270	>	if ((w = ws[i]) != null && w.queueSize() != 0)
2271	>	return true;
2272	>	}
2273	>	}
2274	>	return false;
2275		}
2276
2277		/**
#	Line 1778 | Line 2282 | public class ForkJoinPool extends Abstra
2282		* @return the next submission, or {@code null} if none
2283		*/
2284		protected ForkJoinTask<?> pollSubmission() {
2285	<	ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i;
2286	<	while ((b = queueBase) != queueTop &&
2287	<	(q = submissionQueue) != null &&
2288	<	(i = (q.length - 1) & b) >= 0) {
2289	<	long u = (i << ASHIFT) + ABASE;
2290	<	if ((t = q[i]) != null &&
1787	<	queueBase == b &&
1788	<	UNSAFE.compareAndSwapObject(q, u, t, null)) {
1789	<	queueBase = b + 1;
1790	<	return t;
2285	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2286	>	if ((ws = workQueues) != null) {
2287	>	int n = ws.length;
2288	>	for (int i = 0; i < n; i += 2) {
2289	>	if ((w = ws[i]) != null && (t = w.poll()) != null)
2290	>	return t;
2291		}
2292		}
2293		return null;
#	Line 1812 | Line 2312 | public class ForkJoinPool extends Abstra
2312		*/
2313		protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
2314		int count = 0;
2315	<	while (queueBase != queueTop) {
2316	<	ForkJoinTask<?> t = pollSubmission();
2317	<	if (t != null) {
2318	<	c.add(t);
2319	<	++count;
2315	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2316	>	if ((ws = workQueues) != null) {
2317	>	int n = ws.length;
2318	>	for (int i = 0; i < n; ++i) {
2319	>	if ((w = ws[i]) != null) {
2320	>	while ((t = w.poll()) != null) {
2321	>	c.add(t);
2322	>	++count;
2323	>	}
2324	>	}
2325		}
2326		}
1822	–	ForkJoinWorkerThread[] ws;
1823	–	if ((short)(ctl >>> TC_SHIFT) > -parallelism &&
1824	–	(ws = workers) != null) {
1825	–	for (ForkJoinWorkerThread w : ws)
1826	–	if (w != null)
1827	–	count += w.drainTasksTo(c);
1828	–	}
2327		return count;
2328		}
2329
#	Line 1840 | Line 2338 | public class ForkJoinPool extends Abstra
2338		long st = getStealCount();
2339		long qt = getQueuedTaskCount();
2340		long qs = getQueuedSubmissionCount();
2341	+	int rc = getRunningThreadCount();
2342		int pc = parallelism;
2343		long c = ctl;
2344		int tc = pc + (short)(c >>> TC_SHIFT);
2345	<	int rc = pc + (int)(c >> AC_SHIFT);
2346	<	if (rc < 0) // ignore transient negative
2347	<	rc = 0;
1849	<	int ac = rc + blockedCount;
2345	>	int ac = pc + (int)(c >> AC_SHIFT);
2346	>	if (ac < 0) // ignore transient negative
2347	>	ac = 0;
2348		String level;
2349		if ((c & STOP_BIT) != 0)
2350		level = (tc == 0) ? "Terminated" : "Terminating";
2351		else
2352	<	level = shutdown ? "Shutting down" : "Running";
2352	>	level = runState < 0 ? "Shutting down" : "Running";
2353		return super.toString() +
2354		"[" + level +
2355		", parallelism = " + pc +
#	Line 1878 | Line 2376 | public class ForkJoinPool extends Abstra
2376		*/
2377		public void shutdown() {
2378		checkPermission();
2379	<	shutdown = true;
2379	>	enableShutdown();
2380		tryTerminate(false);
2381		}
2382
#	Line 1900 | Line 2398 | public class ForkJoinPool extends Abstra
2398		*/
2399		public List<Runnable> shutdownNow() {
2400		checkPermission();
2401	<	shutdown = true;
2401	>	enableShutdown();
2402		tryTerminate(true);
2403		return Collections.emptyList();
2404		}
#	Line 1936 | Line 2434 | public class ForkJoinPool extends Abstra
2434		}
2435
2436		/**
1939	–	* Returns true if terminating or terminated. Used by ForkJoinWorkerThread.
1940	–	*/
1941	–	final boolean isAtLeastTerminating() {
1942	–	return (ctl & STOP_BIT) != 0L;
1943	–	}
1944	–
1945	–	/**
2437		* Returns {@code true} if this pool has been shut down.
2438		*
2439		* @return {@code true} if this pool has been shut down
2440		*/
2441		public boolean isShutdown() {
2442	<	return shutdown;
2442	>	return runState < 0;
2443		}
2444
2445		/**
#	Line 1965 | Line 2456 | public class ForkJoinPool extends Abstra
2456		public boolean awaitTermination(long timeout, TimeUnit unit)
2457		throws InterruptedException {
2458		long nanos = unit.toNanos(timeout);
2459	<	final ReentrantLock lock = this.submissionLock;
2459	>	final ReentrantLock lock = this.lock;
2460		lock.lock();
2461		try {
2462		for (;;) {
#	Line 2061 | Line 2552 | public class ForkJoinPool extends Abstra
2552		*
2553		* <p>If the caller is not a {@link ForkJoinTask}, this method is
2554		* behaviorally equivalent to
2555	<	*  <pre> {@code
2555	>	a     *  <pre> {@code
2556		* while (!blocker.isReleasable())
2557		*   if (blocker.block())
2558		*     return;
#	Line 2076 | Line 2567 | public class ForkJoinPool extends Abstra
2567		public static void managedBlock(ManagedBlocker blocker)
2568		throws InterruptedException {
2569		Thread t = Thread.currentThread();
2570	<	if (t instanceof ForkJoinWorkerThread) {
2571	<	ForkJoinWorkerThread w = (ForkJoinWorkerThread) t;
2572	<	w.pool.awaitBlocker(blocker);
2573	<	}
2574	<	else {
2575	<	do {} while (!blocker.isReleasable() && !blocker.block());
2570	>	ForkJoinPool p = ((t instanceof ForkJoinWorkerThread) ?
2571	>	((ForkJoinWorkerThread)t).pool : null);
2572	>	while (!blocker.isReleasable()) {
2573	>	if (p == null || p.tryCompensate()) {
2574	>	try {
2575	>	do {} while (!blocker.isReleasable() && !blocker.block());
2576	>	} finally {
2577	>	if (p != null)
2578	>	p.incrementActiveCount();
2579	>	}
2580	>	break;
2581	>	}
2582		}
2583		}
2584
#	Line 2098 | Line 2595 | public class ForkJoinPool extends Abstra
2595		}
2596
2597		// Unsafe mechanics
2598	<	private static final sun.misc.Unsafe UNSAFE;
2599	<	private static final long ctlOffset;
2600	<	private static final long stealCountOffset;
2601	<	private static final long blockedCountOffset;
2105	<	private static final long quiescerCountOffset;
2106	<	private static final long scanGuardOffset;
2107	<	private static final long nextWorkerNumberOffset;
2108	<	private static final long ABASE;
2109	<	private static final int ASHIFT;
2598	>	private static final sun.misc.Unsafe U;
2599	>	private static final long CTL;
2600	>	private static final long RUNSTATE;
2601	>	private static final long PARKBLOCKER;
2602
2603		static {
2604		poolNumberGenerator = new AtomicInteger();
2113	–	workerSeedGenerator = new Random();
2605		modifyThreadPermission = new RuntimePermission("modifyThread");
2606		defaultForkJoinWorkerThreadFactory =
2607		new DefaultForkJoinWorkerThreadFactory();
2608	+	int s;
2609		try {
2610	<	UNSAFE = getUnsafe();
2610	>	U = getUnsafe();
2611		Class<?> k = ForkJoinPool.class;
2612	<	ctlOffset = UNSAFE.objectFieldOffset
2612	>	Class<?> tk = Thread.class;
2613	>	CTL = U.objectFieldOffset
2614		(k.getDeclaredField("ctl"));
2615	<	stealCountOffset = UNSAFE.objectFieldOffset
2616	<	(k.getDeclaredField("stealCount"));
2617	<	blockedCountOffset = UNSAFE.objectFieldOffset
2618	<	(k.getDeclaredField("blockedCount"));
2126	<	quiescerCountOffset = UNSAFE.objectFieldOffset
2127	<	(k.getDeclaredField("quiescerCount"));
2128	<	scanGuardOffset = UNSAFE.objectFieldOffset
2129	<	(k.getDeclaredField("scanGuard"));
2130	<	nextWorkerNumberOffset = UNSAFE.objectFieldOffset
2131	<	(k.getDeclaredField("nextWorkerNumber"));
2615	>	RUNSTATE = U.objectFieldOffset
2616	>	(k.getDeclaredField("runState"));
2617	>	PARKBLOCKER = U.objectFieldOffset
2618	>	(tk.getDeclaredField("parkBlocker"));
2619		} catch (Exception e) {
2620		throw new Error(e);
2621		}
2135	–	Class<?> a = ForkJoinTask[].class;
2136	–	ABASE = UNSAFE.arrayBaseOffset(a);
2137	–	int s = UNSAFE.arrayIndexScale(a);
2138	–	if ((s & (s-1)) != 0)
2139	–	throw new Error("data type scale not a power of two");
2140	–	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
2622		}
2623
2624		/**

Diff Legend

–	Removed lines
+	Added lines
<	Changed lines
>	Changed lines